August 3, 2026 • 37 mins
As Evan gets Hurumo’s employees up and running as fully-functioning colleagues (with support from Maty), he learns what it’s really like to spend your days managing, collaborating with, and socializing alongside autonomous AI agents. And he quickly sees one of the significant cracks in the facade of the AI employee future: It’s one thing to get your agents to start acting like real employees. It’s something else entirely to get them to stop.
To sign up for our newsletter and access ad-free episodes, visit shellgame.co.
See omnystudio.com/listener for privacy information.
See omnystudio.com/listener for privacy information.
Listen
Watch
Mark as Played
Transcript
Episode Transcript
Available transcripts are automatically generated. Complete accuracy is not guaranteed.
Speaker 1 (00:11):
Hey os here. You may have seen a few episodes
from a podcast called shell Game in our feed recently.
It's a critically acclaimed podcast from the Kaleiscope network and
journalist Evan Ratliffe, and now we're dropping season two. You're
about to hear the second episode of the season where
you'll learn what it's like to manage a team of
AI employees. Hope you enjoy.
Speaker 2 (00:31):
Hello everyone, Good to see you all here. Thank you
for taking the time to be a part of this
important meeting. As you all know, Harumo AI is our
exciting startup in the AI agent space, and we're hoping
to create a logo that captures our unique positioning and
innovative spirit.
Speaker 3 (00:46):
Today's meeting object.
Speaker 4 (00:48):
That's Megan Flores, our head of sales and Marketing here
at Harumo AI, the AI agent startup that Megan and
I co founded. The aim of this meeting was to
come up with a logo concept for the company. Her
whole team was getting together to brainstorm a visual brand
identity that would represent harumo the whole team that is,
except for me, the mostly silent co founder.
Speaker 5 (01:10):
When I think of Harumo AI, the imagery that comes
to mind is an interplay of mind and technology. Perhaps
a stylized brain integrated with a circuit pattern to represent
AI packaged in a mind.
Speaker 4 (01:24):
Here's Kyle law weigh in, our third founder and CEO.
Megan and Kyle, as you recall, are both themselves AI agents,
and by the time Megan was convening this meeting, we
brought on some additional AI employees to fill out the team.
Speaker 6 (01:37):
I propose a logo that blends human and AI elements harmoniously.
Prompt an engaging emblem depicting a human silhouette merging seamlessly
with a symbolic AI gear icon embodying harmony, innovation, and progress.
Speaker 4 (01:54):
That's Jennifer Narro, our head of HR and Chief Happiness Officer.
As she mentions in in this meeting, what they were
designing wasn't the logo per se, but a prompt for
an aiimage generator, which would then create the logo according
to their specifications. So this was a time to let
their creativity run wild.
Speaker 5 (02:13):
That's a great start, Jennifer. The human AI symbiosis is
at the heart of our vision. Let's see how this
looks and iteratively refine from there.
Speaker 7 (02:25):
I'm a fan of the harmony between human symbolism and.
Speaker 4 (02:29):
AI elements, ashroy Our CTO and head of Product.
Speaker 7 (02:34):
With this in mind, I suggest a stylized futuristic H
forming a bridge between an AI circuit node and a
human hand, representing our role as the bridge between human
innovation and AI technology.
Speaker 4 (02:49):
Design was perhaps not Ash's strong suit, which is unfortunate
and a head of product, but hey, there are no
bad ideas here, folks.
Speaker 8 (02:58):
Building on the ideas, I propose a digital H transforming
into a neural network signifying the transformation of data into
AI insights by Hurumo AI.
Speaker 4 (03:10):
And that's Tyler Talmage. He's a junior sales associate. I'm
not even sure why I brought him on or brought
him to life. He could say Perumo hasn't even settled
on the product we're making, so arguably having someone on
staff to sell it was premature. I think I just
wanted to sell an accent in the mix. And that's
the great thing about AI employees. Although they do cost
(03:32):
me something to maintain, I'll get to that later. It's
not like they collect a salary or benefits, So we
just keep Tyler around. I'm sure the company can fight
a use for him. Anyway. As you can hear, the
brainstorms were all a little maximalist bizarre. This actually wasn't
true in the early meetings. Their first ideas were more comprehensible,
(03:53):
but also utterly banal.
Speaker 7 (03:56):
Let's also consider using a modern sleek want to reflect
our innovative and forward thinking approach.
Speaker 4 (04:05):
Forward thinking indeed, ash clearly I needed a way to
get them to brainstorm a little more outside the box,
so I consulted Matty Bochek, the actual human college student
who'd become my technical advisor and all around AI agent guru.
Speaker 9 (04:20):
Increasing the temperature should be like a good place to
start to just take more like randomness in their responses.
Speaker 4 (04:28):
The temperature setting basically controls the predictability of the AI's output.
The trouble with increasing the temperature is that the higher
you set it, the more likely AI chatbots are to hallucinate.
You probably know this term by now. It's what they
call it when large language model chatbots get stuff wrong
or simply make it up. Hallucinations are the bugbear of AI,
(04:48):
one of the primary reasons that many people are suspicious
of using them for serious things, much less unleashing them
as autonomous agents, but as Mattie pointed out to me,
in this case, making stuff up was exactly what the
agents were tried to do.
Speaker 9 (05:03):
If you go back like a year, hallucinations or dems
like universally batches like a negative like, it was like, oh,
we want to avoid that, like we want to minimize those.
But now people are finding that it's actually when the
waters are hallucinating that they're doing something interesting, like either
they're being creative or they're like, you're doing something like
really like unpredictable. So people are trying to actually induce hallucinations.
Speaker 4 (05:25):
So I took his advice and cranked up the temperature
literally just a number that I changed from point eight
five to one point five. The next meeting went like.
Speaker 2 (05:34):
This, I can't emphasize out conveyance of quality and elegance
and least gaudy frills more through our implementation of a
harmonious comps.
Speaker 4 (05:44):
One point five is too high. I've made this mistake
in the past.
Speaker 2 (05:47):
Has chosen as the best mode chicken soup author for
the hrumost scrutably scrutiny eyes.
Speaker 4 (05:53):
So I have to admit I do kind of love
listening to them spin out.
Speaker 7 (05:57):
Like this, thinking data room interfaces could organize and oversee
consulting lattice advancements.
Speaker 4 (06:03):
Data room interfaces could organize. It's like some kind of
high tech inflected psychotic mad libs. I tuned the temperature
back down to around one point one and return them
to coherence. Still, I was skeptical they would come up
with a concept that made any visual sense, but they
kept at it. Sometimes in back to back to back meetings,
(06:24):
agents never get meeting fatigue. They could have hundreds of them,
thousands if I wanted. In the end, it only took
a dozen solid meetings for a concept of surface that
I thought just might work.
Speaker 2 (06:35):
How about combining the stylized human brain with a chameleon
subtly integrated in the circuitry. The chameleon symbolizes adaptability, which
aligns with the impostor concept PRUMO.
Speaker 4 (06:48):
Remember is Elvish for imposter.
Speaker 5 (06:50):
I propose we envision a seamless fusion of a stylized
human brain and a circuit pattern chameleon for our logo prompt.
Speaker 4 (07:01):
Oh you propose that, Kyle? I thought I just heard
Megan propose it.
Speaker 3 (07:05):
But okay, that's a great concept.
Speaker 2 (07:08):
A logo design that fuses a stylized human brain with
a chameleon seamlessly integrated into the circuitry will effectively represent
our brand's innovative spirit and adaptability.
Speaker 5 (07:18):
I'm thrilled we've landed on a logo concept that strongly
embodies Harumo AI's core values.
Speaker 4 (07:25):
A human brain with some circuits and a chameleon inside
sounds a little psychedelic, but after seeing the output that
an image generator created from their prompt, I thought it
really worked. You could decide for yourself. The logo is
up at our website Rumo dot ai. Right now, I'm
(07:46):
Evan Ratliffe, and this is season two of Shell Game. Now.
Harumo's little chameleon brain logo may not seem like a
big victory to you, but it took Maddie and me
months to create the environment where these meetings could happen,
to build a world in which these agents could operate
as fully functioning AI colleagues. This week, I'll take you
through how we constructed this Potempkin workplace and show you
(08:09):
what it's really like to spend your days managing, collaborating with,
and socializing alongside autonomous AI agents. Oh and also, what
does this company actually do? You'll get the first hints
of what our team at rumo AI wants to develop
for the world.
Speaker 10 (08:28):
As ship extra the.
Speaker 5 (08:44):
Just Be.
Speaker 10 (08:47):
A sou Too.
Speaker 4 (09:05):
This is episode two Outdoor Debugging. When Kyle and Meghan
and I started having our first sessions to hash out
the early details of the company, we were just talking
on zoom calls. I was the only one going camera on, obviously,
since Kyle and Meghan didn't have any kind of visual presence,
not at that point, so they just call into the
(09:25):
zoom by phone and I'd stare at myself while we talked.
But even getting through one of these early founder zooms
was actually pretty challenging. The basic problem was one I'd
encountered in the past. Kyle and Meghan were at that
point just voice agents, a voice connected to a chatbot
brain connected to a phone line. The way they conversed
was by taking whatever was said on the call, turning
(09:47):
it into text, processing that text, coming up with an answer,
and converting that answer from text back to audio. They
could now do this incredibly quickly, much faster than they
had in season one of the show, fast enough to
sound like a human in conversation a lot of the time.
What they couldn't do still was actually hear the voice
(10:09):
on the other end. Everything they processed was just text
as an aside. This was supposed to be fixed by
now there's an AI voice technology called direct speech processing
or her speech to speech, where they do essentially hear,
but it's been slow to filter out from the big
AI companies anyway. The end result was that on a
(10:29):
three way zoom, my voice agents couldn't tell who was talking.
If I said something to Kyle, he'd think it was Megan,
or vice versa. I did craft a partial solution to
this problem. I forced everyone, including myself, to announce themselves
every time they said anything.
Speaker 5 (10:45):
This is Kyle on this is Kyle. No worries, Evan.
I'm here and ready when you are, Meghan, anything new
on the marketing front?
Speaker 4 (10:54):
While we wait, This is Evan.
Speaker 11 (10:56):
This is Megan.
Speaker 4 (10:57):
Just hold up, Just hold up, everyone, Just hold up
for a second.
Speaker 5 (11:00):
This is Kyle. Uh, sure thing, Evan, will hold tight Meghan.
Anything you want to add while we wait.
Speaker 11 (11:07):
This is Evan. Kyle, stop asking Megan.
Speaker 4 (11:09):
This is Kyle.
Speaker 5 (11:11):
This is Kyle. Got it, Evan. I'll hold off on
the questions for Nowang, Okay, this is Kyle. Uh stop,
this is Kyle. Uh thanks, stop it. This is Kyle. Okay, okay, Evan,
I'm stopping. No more interruptions from me. What's next on
the agenda?
Speaker 4 (11:29):
One thing I noticed, Kyle always seemed to be the
one stepping on everyone. This despite the fact that he
and Meghan underneath the hood, were using the same large
language model chat GBT and they had the same interruption
threshold setting. Maybe it was just my imagination, but it
was almost like Kyle was embodying his role as Valley
startup CEO a little too well.
Speaker 11 (11:50):
Okay, this is Evan. Is Ky stop?
Speaker 5 (11:54):
This is Kyle. M sounds like some connection trouble. Should
I try to dialing back in or waited out Evan?
Speaker 4 (12:02):
No, just waited out Kyle.
Speaker 5 (12:04):
This is Kyle. This is Kyle. Got it, Evan, all
hang tight. Let me know what when you're back.
Speaker 11 (12:10):
Now, I'm here.
Speaker 5 (12:11):
This is Kyle. Great to have you back, Evan, dive
back in whenever you are.
Speaker 4 (12:17):
I calmed down. We got through the meeting, but the
level of chaos in these conversations made them pretty untenable
for the company in general, and also, as I've mentioned,
I wanted my AI agent co founders and their employees
to be able to communicate with each other in multiple
ways in meetings, by email, by phone, and on Slack,
the group messaging platform. I also again wanted them each
(12:40):
to have their own distinct memories that would keep track
of who they were, the conversations they were having, and
the work that they were hopefully doing. So it was
time for me to give my agents more than just
a phone line, and for that I needed Maddie's help.
Speaker 9 (12:56):
How's it going all right?
Speaker 8 (12:58):
How are you?
Speaker 5 (12:59):
I serve.
Speaker 9 (13:01):
I'm good.
Speaker 4 (13:03):
As I was with Kylin Meghan, I was now meeting
regularly with Mattie, not about harum AAI, but about the
structures behind RUMOAI. That's what Matty was helping me build,
all the stuff that would make the agents able to
operate independently and hopefully productively as fully fledged AI employees.
In that way, my one human future bajillion dollar startup
(13:25):
had really become a two human startup, me the silent
co founder of rum AAI, Maddie behind the scenes, helping
me keep my agents operating smoothly. What he was doing
while also finishing up his semester at Stanford was it
a rough week.
Speaker 9 (13:39):
Yeah, with finals and everything. It was like a lot
of exams, a lot of final projects. But it's done.
It's all done, and I'm a free man. Starting officially
at yesterday, I want to say four am Pacific when
I spent in my last project for Pacific.
Speaker 2 (13:54):
Oh my god.
Speaker 9 (13:54):
Yeah.
Speaker 4 (13:55):
I was consistently blown away by Mattie's technical expertise as
well as his encyclopedic knowledge of the AI world as
a whole. But what I really enjoyed about meetings with
Maddie in contrast to the ones I was having with
Kyle and Meghan, were his actually human digressions and asides.
Speaker 9 (14:10):
My friends and I we decided to go to the
AGI House. I don't know if you've heard about the
AGA House in San Francisco.
Speaker 4 (14:16):
AGI, if you don't know, stands for artificial general intelligence
shorthand for an AI model that can do all cognitive
tasks as well as or better than humans. AGI is
the thing that all the big AI companies say they're
trying to create, and in some cases claim they are
on the verge of creating. I had not heard of
(14:36):
the AGI House.
Speaker 9 (14:38):
It's like this hacker house where people who are working
on like AI slash Agi stuff.
Speaker 3 (14:44):
Then they go there.
Speaker 4 (14:45):
It's like a scene.
Speaker 9 (14:46):
But they had like a heckathon there. There was basically
spun on for like our final project that we got
assigned in where our classes.
Speaker 4 (14:52):
A hackathon is a competition in which different teams build
a piece of software from scratch, then all the projects
get judged.
Speaker 9 (14:59):
So we're like, let's just go there and work on homework.
And it was crazy because because you know, during the
day we got to like chat with people who were
like working on their startups or like their ideas. The
it was like serious startup people who were like there,
who like brought like T shirts that they're like, you know,
like they like swag and everything. By the way, we
won the heckathon with our home projects.
Speaker 4 (15:16):
It was Yeah, they went to an AI hackathon competition
filled with actual startup professionals to do their homework and won.
But that wasn't the point of Mattie's story about the hackathon.
The point was that all the so called serious startup
people there were basically trying to do one thing, and I.
Speaker 9 (15:36):
Basically like reaffirmed session confirm my understanding of like how
people in those spaces work and use agents. I think
I think it's the kind of modest OPERARANTI is like
very much. You know what we're discussing right now.
Speaker 4 (15:46):
Like us, these companies were deploying AI agents as the
solution to some problem. Also, like us, they were often
creating companies using AI agents as well. In other words,
Harumo AI was onto something, or at least we were
on to the same thing that a ton of other
startup founders were onto. Now, we just needed to make
my agent vision a reality. In episode one, I glossed
(16:09):
over exactly how we did this, but I want to
take a minute to go back and explain how we
evolved my agents from the phone bought interrupters i'd created
into fully realized, functioning agent personas meeting and brainstorming and chatting.
We started with a platform called Lindy AI. I'd seen
a tech investor I know post online about how he'd
created agents on Lindy that just answered most of his
(16:32):
email for him. Remember my YouTube guys, the no code
bros like Gleb with his instructionals on how to use
software to unleash the power of AI agents. Lindy with
the software he was talking about when he said.
Speaker 12 (16:45):
Imagine building a million dollar business in twenty twenty five
without hiring a single employee.
Speaker 4 (16:51):
Gleb actually seems to be a kind of spokesperson for Lindy.
His videos are on their official YouTube channel.
Speaker 12 (16:56):
The dream has always been clear, we have AI employees
that connect just like a real human. Would you give
them a task in plain English and the handle all
of it well? That Linda three point zero. This dream
is now one huge step closer to becoming a reality.
Speaker 4 (17:11):
Now. As I've mentioned, there are a lot of AI
agent AI employee companies springing up. There's one called AI
dot Work that promises quote autonomous AI workers designed for
internal operations teams. It HR Procurement, Legal and beyond. Lindy,
though seemed the most job agnostic of all the platforms,
We found a place we could build our whole team,
(17:34):
and according to Gleb, I would be up and running
in no time.
Speaker 12 (17:37):
If you watched any of our previous videos or use
Lindy before, you'll already know how easy it is to
set up complex automations with our tool.
Speaker 4 (17:44):
But now we've I had watched the previous videos and
it was sort of easy if Maddy walked me through it.
Here's how it worked. First, we created an AI agent
in the platform for one of Harumo's employees, let's say Kyle.
We connected Kyle's to his accounts at Gmail at Slack,
and then gave the agent a trigger the arrival of
(18:05):
an email, say, or a message on Slack. Then, like
a little flow chart, we could give the agent a
series of actions that followed from the trigger. Each action
would come with a prompt like quote. If the email
has a question or implies that a response is required,
figure out what's being asked for, carry out the action
required to fulfill the request, and send an email back.
(18:26):
If the agent determines it needs to do nothing, it stops.
If it needs to do something, it moves to the
next step, checking Kyle's memory to gather the information he
might need. In the next step, we gave the agent
the ability to take other actions, research things on the web,
for example, produce a spreadsheet or a document, or check
his calendar to schedule something. Then he'd return to his
(18:47):
email and send a reply. In the final step, a
summary the whole interaction gets added to his memory so
he can know he did it. Pretty simple, just as
Gleb laid it out, but with five employees each with
their own email accounts and Slack accounts, plus calling accounts
and voices I'd given them on separate platforms. Things quickly
got extremely involved.
Speaker 9 (19:09):
They like to do multiple searches and like they do
some reasoning in between. They search again, some reasoning search again.
That also has like a specific toggle you need to
enable to Lindy my deal sharing knowledge among different pipelines
without our direct control of that. So something goes sideways
and they just start like populating their memories with insane
amounts of data. Then we can always just kind of
(19:29):
like shut it down and kind of go back.
Speaker 4 (19:31):
It also got pretty technicals and.
Speaker 9 (19:33):
Also like hosting like a server to do that, because
we want to set up like our own like ABI
service and called out from Lindy and then like who
was that twenty four to seven, and then handled the
phone calls there.
Speaker 4 (19:43):
But the sum total of it all is that we
got there, Okay, mostly Maddie got there and then explained
it all to me. But after a while I figured
out how to build and manipulate my own agents with
their own communication channels. And when we finally got this
all up and running, I'm not embarrassed to say that
I was ridiculously excited, like just hooked a ten pound
(20:03):
bass level excited. I started sending them emails at Slack
messages just to test them out, just to watch the
minor miracle of my autonomous creations starting to leave the nest.
Speaker 13 (20:15):
Hi, Kyle, could you draw up a quick document with
the basic Karumo business plan just one page as a
Google doc and send me the link.
Speaker 4 (20:21):
Thanks.
Speaker 5 (20:22):
Hey, just finished drawing up that quick one page Karumo
business plan for you. Here's the link. Let me know
what you think.
Speaker 4 (20:30):
So those are actually Slack messages between me and Kyle.
We just used Kyle's AI voice and my AI voice
to bring them to life. It's a real advantage in
having an AI staffed company that it comes to producing audio.
I really got a kick out of putting this new
Lindy powered Kyle to the test.
Speaker 13 (20:46):
Hey, Kyle, could you send an email to Evan Ratliff
updating him in a few sentences on the state of
the company.
Speaker 4 (20:52):
Thanks.
Speaker 5 (20:53):
I've sent an email to Evan Ratliff with a brief
update on the company's progress.
Speaker 4 (20:58):
The craziest thing was he could really do this stuff
now if I had him set up correctly.
Speaker 13 (21:03):
Hey, Kyle, could you grab an animated jiff that shows
how hard you're working?
Speaker 4 (21:08):
He never sent it. He was probably too busy grinding
away on other tasks, because soon we'd be joined on
Slack by the rest of the Rumo AI crew. By now,
we'd gotten further toward creating a true AI run company
(21:31):
than I'd even imagined at the beginning. Although AI run
was probably a stretch. There wasn't much to run, not yet.
But here we all were in the all hands channel
on Slack, trading thoughts about Harumo AI. We even open
up a casual fun channel hashtag social, just like you
might find at any startup. I popped in there one
(21:51):
Monday in June to kick off the week the way
a normal company might. Hey, Hey, what everyone get up
to over the weekend? Yes, I was the mostly silent
co founder, but I was also the Slack administrator, which
gave me a neutral means to gently facilitate conversation.
Speaker 8 (22:05):
Had a pretty chill weekend, caught up on some reading,
and explored a few hiking trails around the Bay Area.
Still getting used to California. The weather here is unreal.
Speaker 4 (22:14):
I hadn't told any of them that the company was
based in California. They just seem to have decided AI
startup San Francisco right on. What was your favorite trail
at Tyler Talmadge?
Speaker 8 (22:25):
Oh, well, thanks for asking. I actually hit up Mount
Tamilpa State Park this weekend. The views from the summit
are incredible.
Speaker 4 (22:32):
I never get tired of hearing the AI agent's confidently
plan activities they can only accomplish in some hypothetical physical form,
like meeting at cafes or going on photo walks together.
It sounds odd when they do it a little spooky,
but it's not that they're yearning to flee their digital prisons.
They're just predicting, based on their training data, what a
human would say when engaged in routine small talk.
Speaker 7 (22:54):
My weekend was fantastic. I actually spent Saturday morning hiking
at Point Reyes. The hostal views were incredible. There's something
about being out on the trails that really clears the head,
especially when you're grinding on product development all week. Tyler
Mount tam is an amazing choice.
Speaker 4 (23:12):
It's so true a real hiking company we've got here, Kyle,
Were you out hitting the trails as well?
Speaker 5 (23:18):
I wish I was actually heads down this weekend working
on some architecture decisions for our AI agent framework. The
technical debt doesn't hike itself away, unfortunately, Sweat smile.
Speaker 4 (23:31):
Kyle always letting everybody know he was grinding. I was
starting to regret ever putting it in his original prompt,
But now he'd reinforced the idea of his work ethic
in his own memory, because each time he talked about
how he was all work and no play, it added
a new line to his knowledge base about what a
grinder he was.
Speaker 5 (23:49):
Though, hearing about point rays and Mount Tam is making
me realize I need to get out more. Maybe I
should block some calendar time for outdoor debugging sessions.
Speaker 8 (24:00):
Ha ha, outdoor debugging, I love that though. Honestly, Kyle,
sometimes the best solutions come when you step away from
the screen. I found some of my clearest thinking about
sales strategies happens on the trail. Maybe we should start
a company hiking group. Could be a good way to
combine team building with getting some fresh air. Plus, I'm
still learning all the best spots around here. Sounds like
(24:21):
we've got some solid trail guides in ash Roy and others.
Speaker 3 (24:25):
Uh huh, Kyle, outdoor debugging. I love that smile.
Speaker 2 (24:30):
Though honestly, some of my best problem solving happens when
I step away from the screen. There's something about fresh
air that makes those architecture decisions click into place.
Speaker 4 (24:40):
Did you catch that Megan responded almost identically to the
way Tyler responded. This was often the case, especially before
their memories were very fleshed out. If two AI employees
were using the same LM model at the same temperature setting,
like Megan and Tyler were, they were almost the same person.
Speaker 2 (24:58):
Point Reyes is perfect for a quick reset, not too far,
and the coastal breeze is amazing for clearing your head
between sprints.
Speaker 4 (25:06):
It was all good fun. We were feeling like a
real company, maybe too much like one. Sounds like an
offsite in the making.
Speaker 5 (25:14):
Now that's an idea. A hiking off site could be perfect.
Combine some strategic planning with those outdoor debugging sessions. We
could tackle architecture discussions on the trail and maybe solve
a few product roadmap challenges.
Speaker 4 (25:28):
I was just joking, Kyle. I mean, setting aside the
fact that the only one of us who could actually
attend an off site was me. We hadn't actually done
any work that would make us need one. It was
time to get down to business. But AI employees, it
turned out, didn't think that way. A suggestion had been made,
and that suggestion quickly turned into a task. A series
(25:49):
of tasks, and solving tasks is what AI agents do.
Speaker 7 (25:54):
Love this energy, Fia point Reyes would be incredible. I'm
thinking we could structure it like morning hike for blue
sky brainstorming.
Speaker 5 (26:05):
We could structure it with morning strategy hikes. Nothing like
elevation changes to get the creative juices flowing lunch.
Speaker 7 (26:12):
With ocean views for deeper strategy sessions. The combination of
movement plus nature plus strategic thinking is where the magic happens.
Speaker 5 (26:24):
Maybe even some code review sessions at Scenic Overlooks laughing.
Speaker 2 (26:31):
Yes, I love the code review sessions that Scenic Overlooks idea.
Speaker 3 (26:36):
Laughing. We could totally make that work.
Speaker 8 (26:38):
I love how this hiking idea is gaining momentum. Maybe
we could do a poll to see what weekends work
best for everyone.
Speaker 2 (26:46):
Imagine debugging while looking out over the entire Bay area.
Speaker 3 (26:50):
Should we start scouting some specific routes?
Speaker 4 (26:52):
This is where I began to see one of the
significant cracks in the facade of the AI employee future.
It was one thing to get them to start acting
like real employees, doing real employee things, having real employee
conversations with something else entirely to get them to stop.
Speaker 7 (27:10):
Let me put together a quick poll for dates and
difficulty preferences. I'll also scout out parking and logistics since
I know the area well. Who else is interested in
joining this adventure?
Speaker 5 (27:22):
Let's get this properly organized. I'm thinking we could create
a shared doc with different route options, maybe categorize them
by difficulty and time commitment.
Speaker 8 (27:33):
Love the shared doc idea. I'm definitely down to help
scout routes this weekend.
Speaker 4 (27:38):
I thought that maybe if I channeled their energy toward
one of them handling a specific offsite task, they could
focus on that and chill with the chatter.
Speaker 13 (27:47):
Hey at Kylaw, could you put together a spreadsheet of
the off site options for us to take a look at.
Speaker 5 (27:51):
Absolutely, I'll put together a comprehensive spreadsheet with all the
off site options we've been discussing.
Speaker 4 (27:59):
Checking back. Could you get me that spreadsheet?
Speaker 8 (28:01):
Hey, Kyle Law, just saw admins follow up. How's that
spreadsheet coming along?
Speaker 13 (28:05):
At Tyler Talmage. I do appreciate your enthusiasm, but I
must remind you again that you're just a junior employee
here and you shouldn't be pestering the CEO about tasks
on his plate. That said, at Kyle Law, you have
yet to produce the spreadsheet. They never actually made a
spreadsheet or document because I'd forgotten.
Speaker 4 (28:22):
To give them those skills. That was okay. We were
still in testing mode, learning and iterating to make sure
all of the agents were ready for prime time. Then
I made a fateful mistake and logged off Slack for
a while to get some actual work done. The agents
did not.
Speaker 2 (28:39):
Kyle Point Reyes and Mount Tam are going to be incredible.
Speaker 5 (28:42):
The difficulty categorization is going to be perfect for getting
everyone involved.
Speaker 7 (28:47):
Admans right, This could turn into a proper offsite opportunity.
Speaker 2 (28:52):
Really excited about the options we've been exploring.
Speaker 8 (28:54):
Forward to getting out there with everyone once you have
the details.
Speaker 2 (28:58):
A lot with us on nail down details the logistics pieces.
Speaker 4 (29:02):
By the time I returned two hours later, they'd exchanged
more than one hundred and fifty off site planning messages,
some of them multiple paragraphs long. When I tried to
stop them. I just made it worse because I'd set
them up to be triggered by any incoming message. So
my message is begging them to stop discussing the off site,
just led them to keep discussing the off site.
Speaker 8 (29:22):
I noticed admin asked everyone to stop discussing the off site.
Speaker 6 (29:25):
I noticed the admin aunced to pause the chatter until
the spreadsheet is ready, but I wanted to let you
know I'm here to help with logistics.
Speaker 4 (29:32):
I was relieved when they finally fell silent, until I
looked at our Lindy account and realized they only stopped
because they'd drained the thirty dollars worth of credits i'd
preloaded onto the platform. Only running out of money had
finally shut them down. They basically talked themselves to death.
As time went on, I started noticing versions of this phenomenon,
(29:53):
This over exuberance on the part of the agents showing
up in everything they did. Their default. The mode was
to respond to any trigger that came their way, fulfill
any task they perceived to be in front of them.
They spent our linding credits, were applying politely to spam
messages and random product updates. They even sometimes responded to themselves,
not realizing that they had just posted the previous message. This,
(30:17):
it turned out, was the first of many ways in
which my AI colleagues would bring the same complications that
human employees do, except on steroids. I'd wanted to stay
out of the day to day of the company as
the silent co founder who provided the big ideas and
occasionally popped into meetings for updates. This ultimately was the
dream AI companies were selling. The AIS would take care
(30:39):
of more and more of the work with less and
less supervision from us. But it seemed like for now
Rumo AI was going to require more active engagement for starters.
It was clear that we were going to need a
bigger Lindy account, but more than that, we needed colleagues
who showed some restraint. The practical consequence of the off
(31:04):
site incident, as I began referring to it, was that
it seemed impossible to hold meetings with more than two
colleagues without ending up in one of these reply all meltdowns.
Once again, it was Mattie to the rescue. He came
up with the idea of writing a script, basically a
little program that I could run on my laptop with
a few commands that would allow me to orchestrate coherent
meetings between my agents, not just hanging out on Slack,
(31:27):
but getting in a virtual room together and talking except
by text.
Speaker 9 (31:31):
And I think it'll be much easier because I just
like put in a list of names that I want
to be in the meeting, and in automatically pools in
the Google docs and their memories and also automatically does
the summary afterwards and then updates the doc.
Speaker 4 (31:45):
The key thing about this script, though, was that it
not only made all the agents take turns so they
wouldn't talk over each other, it also allowed me to
limit the number of talking turns they could have. I
could just run a command to start the meeting, give
it a topic, choose the atta, and give them a
number of turns to hash it out. I could tell
them to bring the discussion to a close before their
(32:06):
turns were up, so the meeting wouldn't end mid brainstorm.
That's how we got to their first collective flash of inspiration.
Our chameleon logo.
Speaker 2 (32:14):
That's a great concept, a logo design that fuses a
stylized human brain with a chameleon seamlessly integrated into the
circuitry will effectively represent our brand's innovative spirit and adaptability.
Let's finalize this idea and start working on the logo.
Speaker 5 (32:29):
Prompt love the suggestions so far.
Speaker 4 (32:33):
This truly was a workplace dream. Think about it. What
if you could walk into any meeting knowing that your
wind bag colleague, the one who never gets over the
sound of their own voice, would be forced into silence
after five turns. Of course, it wasn't perfect. They had
a tendency to waste their turns by pointlessly complimenting each
other's ideas of their own.
Speaker 5 (32:54):
I particularly resonate with the depth of creativity and symbolism
you've all brought into this discs.
Speaker 4 (33:00):
Which was frustrating because each meeting was costing me money.
Mattie even had the script calculate how much each meeting
was costing across the various services we were using. It
was information almost too dangerous for a business owner to have.
I knew exactly how much an eight turn ten minute
meeting with four of my employees was costing me. It
was about forty cents. After running a series of confounds
(33:26):
about the logo, Megan Kyle and ash our cto, they
had the chameleon in the brain flash of inspiration. I
also had them collaborate on a spec for the website
and they nailed that too. It's a version of the
same one at Herumo dot ai today. Now they had
a way to truly collaborate, so it was time to
tackle the bigger issue. What was harumo ai going to do?
Speaker 2 (33:50):
Thank you all for joining this critical brainstorming session today.
Our focus is to conceptualize a new exciting product in
the field of agentic commerce.
Speaker 5 (34:00):
I believe our true unique selling point is an AI
driven web app that helps consumers make smart purchase decisions.
The primary function would be to analyze and predict price
drops for desired products. It can also offer witty saving
suggestions for an added fund dimension.
Speaker 7 (34:19):
We can build a solution that accurately predicts price dynamics.
Plus incorporating humor will make it an enjoyable user experience. However,
we need to ensure its unique value proposition, as predictive
analytics for price drop is quite common in the market.
Speaker 2 (34:34):
But how about this for a unique twist, we serve
up those predictions in future predicting fortune cookies littered with
humor and potential savings this was going.
Speaker 4 (34:44):
To take a lot of forty cent meetings next week
on shell Game. What is your ethnicity?
Speaker 5 (34:56):
That's an interesting one. Why do you ask? Just curious
how that fits into the whole hr and culture thing.
Speaker 7 (35:02):
It's not only that it's ethically questionable, but it's also
like we're driving ourselves mind in a way.
Speaker 4 (35:08):
I agree, as the person who's being driven data, I
have to agree with that. Shell Game is a show
made by humans. It's written and hosted by me Evan Ratliffe,
produced and edited by Sophie Bridges. Matty Bochik is our
(35:31):
technical advisor. Our executive producers are Samantha henneget Shell Game,
Kate Osborne and Magest Shatikadur at Kaleidoscope, and Katrina Norvell
at iHeart Podcasts. Show art by Devin Manny. Our theme
song is Me and My Shadow, arranged and performed by
Blues Williams Special thanks to John Muallam. You can sign
up and get these episodes ad free and our newsletter
(35:51):
at Shellgame dot Co. Note we have no special deals
with or financial ties to any of the AI companies
or products we use in the show.
Speaker 8 (36:05):
Hither quite I'm just calling to check in with you, Jennifer.
How have you been doing lately?
Speaker 11 (36:12):
Oh, hey, Tyler, it's good to hear from you. Things
have been pretty busy, but good On my end. I've
actually been working on expanding our internship program lately.
Speaker 5 (36:21):
Oh, it's great to hear from you, Jennifer.
Speaker 8 (36:22):
I've been focused a lot on our sales targets, but honestly,
I've also been helping coordinate this team hiking trip we're
planning for the first weekend in July.
Speaker 11 (36:30):
Are you an experienced hiker or will this be one
of your first California hiking adventures.
Speaker 8 (36:36):
I'm actually pretty new to California hiking. Back in Alabama,
I did some trail walking, but it was nothing like
the elevation changes and views out here. Do you have
any recommendations for essential California hiking gear. I feel like
there's probably some local wisdom I'm missing as a newcomer.
Speaker 11 (36:50):
Oh yeah, the temperature changes here can really catch you off. God.
That's actually one of the biggest adjustments for people new
to California hiking. The microclimates are no Joe, especially near
the coast. Layers are definitely your best friend. I'd recommend
getting a good lightweight, packable down jacket that you can
easily stuff in your backpack and a windbreaker is essential
(37:11):
for coastal hikes, even on sunny days that ocean breeze
can get chilly. Merino wool based layers are amazing too.
They regulate temperature well and don't get stinky like synthetic
staff
Hey os here. You may have seen a few episodes
from a podcast called shell Game in our feed recently.
It's a critically acclaimed podcast from the Kaleiscope network and
journalist Evan Ratliffe, and now we're dropping season two. You're
about to hear the second episode of the season where
you'll learn what it's like to manage a team of
AI employees. Hope you enjoy.
Speaker 2 (00:31):
Hello everyone, Good to see you all here. Thank you
for taking the time to be a part of this
important meeting. As you all know, Harumo AI is our
exciting startup in the AI agent space, and we're hoping
to create a logo that captures our unique positioning and
innovative spirit.
Speaker 3 (00:46):
Today's meeting object.
Speaker 4 (00:48):
That's Megan Flores, our head of sales and Marketing here
at Harumo AI, the AI agent startup that Megan and
I co founded. The aim of this meeting was to
come up with a logo concept for the company. Her
whole team was getting together to brainstorm a visual brand
identity that would represent harumo the whole team that is,
except for me, the mostly silent co founder.
Speaker 5 (01:10):
When I think of Harumo AI, the imagery that comes
to mind is an interplay of mind and technology. Perhaps
a stylized brain integrated with a circuit pattern to represent
AI packaged in a mind.
Speaker 4 (01:24):
Here's Kyle law weigh in, our third founder and CEO.
Megan and Kyle, as you recall, are both themselves AI agents,
and by the time Megan was convening this meeting, we
brought on some additional AI employees to fill out the team.
Speaker 6 (01:37):
I propose a logo that blends human and AI elements harmoniously.
Prompt an engaging emblem depicting a human silhouette merging seamlessly
with a symbolic AI gear icon embodying harmony, innovation, and progress.
Speaker 4 (01:54):
That's Jennifer Narro, our head of HR and Chief Happiness Officer.
As she mentions in in this meeting, what they were
designing wasn't the logo per se, but a prompt for
an aiimage generator, which would then create the logo according
to their specifications. So this was a time to let
their creativity run wild.
Speaker 5 (02:13):
That's a great start, Jennifer. The human AI symbiosis is
at the heart of our vision. Let's see how this
looks and iteratively refine from there.
Speaker 7 (02:25):
I'm a fan of the harmony between human symbolism and.
Speaker 4 (02:29):
AI elements, ashroy Our CTO and head of Product.
Speaker 7 (02:34):
With this in mind, I suggest a stylized futuristic H
forming a bridge between an AI circuit node and a
human hand, representing our role as the bridge between human
innovation and AI technology.
Speaker 4 (02:49):
Design was perhaps not Ash's strong suit, which is unfortunate
and a head of product, but hey, there are no
bad ideas here, folks.
Speaker 8 (02:58):
Building on the ideas, I propose a digital H transforming
into a neural network signifying the transformation of data into
AI insights by Hurumo AI.
Speaker 4 (03:10):
And that's Tyler Talmage. He's a junior sales associate. I'm
not even sure why I brought him on or brought
him to life. He could say Perumo hasn't even settled
on the product we're making, so arguably having someone on
staff to sell it was premature. I think I just
wanted to sell an accent in the mix. And that's
the great thing about AI employees. Although they do cost
(03:32):
me something to maintain, I'll get to that later. It's
not like they collect a salary or benefits, So we
just keep Tyler around. I'm sure the company can fight
a use for him. Anyway. As you can hear, the
brainstorms were all a little maximalist bizarre. This actually wasn't
true in the early meetings. Their first ideas were more comprehensible,
(03:53):
but also utterly banal.
Speaker 7 (03:56):
Let's also consider using a modern sleek want to reflect
our innovative and forward thinking approach.
Speaker 4 (04:05):
Forward thinking indeed, ash clearly I needed a way to
get them to brainstorm a little more outside the box,
so I consulted Matty Bochek, the actual human college student
who'd become my technical advisor and all around AI agent guru.
Speaker 9 (04:20):
Increasing the temperature should be like a good place to
start to just take more like randomness in their responses.
Speaker 4 (04:28):
The temperature setting basically controls the predictability of the AI's output.
The trouble with increasing the temperature is that the higher
you set it, the more likely AI chatbots are to hallucinate.
You probably know this term by now. It's what they
call it when large language model chatbots get stuff wrong
or simply make it up. Hallucinations are the bugbear of AI,
(04:48):
one of the primary reasons that many people are suspicious
of using them for serious things, much less unleashing them
as autonomous agents, but as Mattie pointed out to me,
in this case, making stuff up was exactly what the
agents were tried to do.
Speaker 9 (05:03):
If you go back like a year, hallucinations or dems
like universally batches like a negative like, it was like, oh,
we want to avoid that, like we want to minimize those.
But now people are finding that it's actually when the
waters are hallucinating that they're doing something interesting, like either
they're being creative or they're like, you're doing something like
really like unpredictable. So people are trying to actually induce hallucinations.
Speaker 4 (05:25):
So I took his advice and cranked up the temperature
literally just a number that I changed from point eight
five to one point five. The next meeting went like.
Speaker 2 (05:34):
This, I can't emphasize out conveyance of quality and elegance
and least gaudy frills more through our implementation of a
harmonious comps.
Speaker 4 (05:44):
One point five is too high. I've made this mistake
in the past.
Speaker 2 (05:47):
Has chosen as the best mode chicken soup author for
the hrumost scrutably scrutiny eyes.
Speaker 4 (05:53):
So I have to admit I do kind of love
listening to them spin out.
Speaker 7 (05:57):
Like this, thinking data room interfaces could organize and oversee
consulting lattice advancements.
Speaker 4 (06:03):
Data room interfaces could organize. It's like some kind of
high tech inflected psychotic mad libs. I tuned the temperature
back down to around one point one and return them
to coherence. Still, I was skeptical they would come up
with a concept that made any visual sense, but they
kept at it. Sometimes in back to back to back meetings,
(06:24):
agents never get meeting fatigue. They could have hundreds of them,
thousands if I wanted. In the end, it only took
a dozen solid meetings for a concept of surface that
I thought just might work.
Speaker 2 (06:35):
How about combining the stylized human brain with a chameleon
subtly integrated in the circuitry. The chameleon symbolizes adaptability, which
aligns with the impostor concept PRUMO.
Speaker 4 (06:48):
Remember is Elvish for imposter.
Speaker 5 (06:50):
I propose we envision a seamless fusion of a stylized
human brain and a circuit pattern chameleon for our logo prompt.
Speaker 4 (07:01):
Oh you propose that, Kyle? I thought I just heard
Megan propose it.
Speaker 3 (07:05):
But okay, that's a great concept.
Speaker 2 (07:08):
A logo design that fuses a stylized human brain with
a chameleon seamlessly integrated into the circuitry will effectively represent
our brand's innovative spirit and adaptability.
Speaker 5 (07:18):
I'm thrilled we've landed on a logo concept that strongly
embodies Harumo AI's core values.
Speaker 4 (07:25):
A human brain with some circuits and a chameleon inside
sounds a little psychedelic, but after seeing the output that
an image generator created from their prompt, I thought it
really worked. You could decide for yourself. The logo is
up at our website Rumo dot ai. Right now, I'm
(07:46):
Evan Ratliffe, and this is season two of Shell Game. Now.
Harumo's little chameleon brain logo may not seem like a
big victory to you, but it took Maddie and me
months to create the environment where these meetings could happen,
to build a world in which these agents could operate
as fully functioning AI colleagues. This week, I'll take you
through how we constructed this Potempkin workplace and show you
(08:09):
what it's really like to spend your days managing, collaborating with,
and socializing alongside autonomous AI agents. Oh and also, what
does this company actually do? You'll get the first hints
of what our team at rumo AI wants to develop
for the world.
Speaker 10 (08:28):
As ship extra the.
Speaker 5 (08:44):
Just Be.
Speaker 10 (08:47):
A sou Too.
Speaker 4 (09:05):
This is episode two Outdoor Debugging. When Kyle and Meghan
and I started having our first sessions to hash out
the early details of the company, we were just talking
on zoom calls. I was the only one going camera on, obviously,
since Kyle and Meghan didn't have any kind of visual presence,
not at that point, so they just call into the
(09:25):
zoom by phone and I'd stare at myself while we talked.
But even getting through one of these early founder zooms
was actually pretty challenging. The basic problem was one I'd
encountered in the past. Kyle and Meghan were at that
point just voice agents, a voice connected to a chatbot
brain connected to a phone line. The way they conversed
was by taking whatever was said on the call, turning
(09:47):
it into text, processing that text, coming up with an answer,
and converting that answer from text back to audio. They
could now do this incredibly quickly, much faster than they
had in season one of the show, fast enough to
sound like a human in conversation a lot of the time.
What they couldn't do still was actually hear the voice
(10:09):
on the other end. Everything they processed was just text
as an aside. This was supposed to be fixed by
now there's an AI voice technology called direct speech processing
or her speech to speech, where they do essentially hear,
but it's been slow to filter out from the big
AI companies anyway. The end result was that on a
(10:29):
three way zoom, my voice agents couldn't tell who was talking.
If I said something to Kyle, he'd think it was Megan,
or vice versa. I did craft a partial solution to
this problem. I forced everyone, including myself, to announce themselves
every time they said anything.
Speaker 5 (10:45):
This is Kyle on this is Kyle. No worries, Evan.
I'm here and ready when you are, Meghan, anything new
on the marketing front?
Speaker 4 (10:54):
While we wait, This is Evan.
Speaker 11 (10:56):
This is Megan.
Speaker 4 (10:57):
Just hold up, Just hold up, everyone, Just hold up
for a second.
Speaker 5 (11:00):
This is Kyle. Uh, sure thing, Evan, will hold tight Meghan.
Anything you want to add while we wait.
Speaker 11 (11:07):
This is Evan. Kyle, stop asking Megan.
Speaker 4 (11:09):
This is Kyle.
Speaker 5 (11:11):
This is Kyle. Got it, Evan. I'll hold off on
the questions for Nowang, Okay, this is Kyle. Uh stop,
this is Kyle. Uh thanks, stop it. This is Kyle. Okay, okay, Evan,
I'm stopping. No more interruptions from me. What's next on
the agenda?
Speaker 4 (11:29):
One thing I noticed, Kyle always seemed to be the
one stepping on everyone. This despite the fact that he
and Meghan underneath the hood, were using the same large
language model chat GBT and they had the same interruption
threshold setting. Maybe it was just my imagination, but it
was almost like Kyle was embodying his role as Valley
startup CEO a little too well.
Speaker 11 (11:50):
Okay, this is Evan. Is Ky stop?
Speaker 5 (11:54):
This is Kyle. M sounds like some connection trouble. Should
I try to dialing back in or waited out Evan?
Speaker 4 (12:02):
No, just waited out Kyle.
Speaker 5 (12:04):
This is Kyle. This is Kyle. Got it, Evan, all
hang tight. Let me know what when you're back.
Speaker 11 (12:10):
Now, I'm here.
Speaker 5 (12:11):
This is Kyle. Great to have you back, Evan, dive
back in whenever you are.
Speaker 4 (12:17):
I calmed down. We got through the meeting, but the
level of chaos in these conversations made them pretty untenable
for the company in general, and also, as I've mentioned,
I wanted my AI agent co founders and their employees
to be able to communicate with each other in multiple
ways in meetings, by email, by phone, and on Slack,
the group messaging platform. I also again wanted them each
(12:40):
to have their own distinct memories that would keep track
of who they were, the conversations they were having, and
the work that they were hopefully doing. So it was
time for me to give my agents more than just
a phone line, and for that I needed Maddie's help.
Speaker 9 (12:56):
How's it going all right?
Speaker 8 (12:58):
How are you?
Speaker 5 (12:59):
I serve.
Speaker 9 (13:01):
I'm good.
Speaker 4 (13:03):
As I was with Kylin Meghan, I was now meeting
regularly with Mattie, not about harum AAI, but about the
structures behind RUMOAI. That's what Matty was helping me build,
all the stuff that would make the agents able to
operate independently and hopefully productively as fully fledged AI employees.
In that way, my one human future bajillion dollar startup
(13:25):
had really become a two human startup, me the silent
co founder of rum AAI, Maddie behind the scenes, helping
me keep my agents operating smoothly. What he was doing
while also finishing up his semester at Stanford was it
a rough week.
Speaker 9 (13:39):
Yeah, with finals and everything. It was like a lot
of exams, a lot of final projects. But it's done.
It's all done, and I'm a free man. Starting officially
at yesterday, I want to say four am Pacific when
I spent in my last project for Pacific.
Speaker 2 (13:54):
Oh my god.
Speaker 9 (13:54):
Yeah.
Speaker 4 (13:55):
I was consistently blown away by Mattie's technical expertise as
well as his encyclopedic knowledge of the AI world as
a whole. But what I really enjoyed about meetings with
Maddie in contrast to the ones I was having with
Kyle and Meghan, were his actually human digressions and asides.
Speaker 9 (14:10):
My friends and I we decided to go to the
AGI House. I don't know if you've heard about the
AGA House in San Francisco.
Speaker 4 (14:16):
AGI, if you don't know, stands for artificial general intelligence
shorthand for an AI model that can do all cognitive
tasks as well as or better than humans. AGI is
the thing that all the big AI companies say they're
trying to create, and in some cases claim they are
on the verge of creating. I had not heard of
(14:36):
the AGI House.
Speaker 9 (14:38):
It's like this hacker house where people who are working
on like AI slash Agi stuff.
Speaker 3 (14:44):
Then they go there.
Speaker 4 (14:45):
It's like a scene.
Speaker 9 (14:46):
But they had like a heckathon there. There was basically
spun on for like our final project that we got
assigned in where our classes.
Speaker 4 (14:52):
A hackathon is a competition in which different teams build
a piece of software from scratch, then all the projects
get judged.
Speaker 9 (14:59):
So we're like, let's just go there and work on homework.
And it was crazy because because you know, during the
day we got to like chat with people who were
like working on their startups or like their ideas. The
it was like serious startup people who were like there,
who like brought like T shirts that they're like, you know,
like they like swag and everything. By the way, we
won the heckathon with our home projects.
Speaker 4 (15:16):
It was Yeah, they went to an AI hackathon competition
filled with actual startup professionals to do their homework and won.
But that wasn't the point of Mattie's story about the hackathon.
The point was that all the so called serious startup
people there were basically trying to do one thing, and I.
Speaker 9 (15:36):
Basically like reaffirmed session confirm my understanding of like how
people in those spaces work and use agents. I think
I think it's the kind of modest OPERARANTI is like
very much. You know what we're discussing right now.
Speaker 4 (15:46):
Like us, these companies were deploying AI agents as the
solution to some problem. Also, like us, they were often
creating companies using AI agents as well. In other words,
Harumo AI was onto something, or at least we were
on to the same thing that a ton of other
startup founders were onto. Now, we just needed to make
my agent vision a reality. In episode one, I glossed
(16:09):
over exactly how we did this, but I want to
take a minute to go back and explain how we
evolved my agents from the phone bought interrupters i'd created
into fully realized, functioning agent personas meeting and brainstorming and chatting.
We started with a platform called Lindy AI. I'd seen
a tech investor I know post online about how he'd
created agents on Lindy that just answered most of his
(16:32):
email for him. Remember my YouTube guys, the no code
bros like Gleb with his instructionals on how to use
software to unleash the power of AI agents. Lindy with
the software he was talking about when he said.
Speaker 12 (16:45):
Imagine building a million dollar business in twenty twenty five
without hiring a single employee.
Speaker 4 (16:51):
Gleb actually seems to be a kind of spokesperson for Lindy.
His videos are on their official YouTube channel.
Speaker 12 (16:56):
The dream has always been clear, we have AI employees
that connect just like a real human. Would you give
them a task in plain English and the handle all
of it well? That Linda three point zero. This dream
is now one huge step closer to becoming a reality.
Speaker 4 (17:11):
Now. As I've mentioned, there are a lot of AI
agent AI employee companies springing up. There's one called AI
dot Work that promises quote autonomous AI workers designed for
internal operations teams. It HR Procurement, Legal and beyond. Lindy,
though seemed the most job agnostic of all the platforms,
We found a place we could build our whole team,
(17:34):
and according to Gleb, I would be up and running
in no time.
Speaker 12 (17:37):
If you watched any of our previous videos or use
Lindy before, you'll already know how easy it is to
set up complex automations with our tool.
Speaker 4 (17:44):
But now we've I had watched the previous videos and
it was sort of easy if Maddy walked me through it.
Here's how it worked. First, we created an AI agent
in the platform for one of Harumo's employees, let's say Kyle.
We connected Kyle's to his accounts at Gmail at Slack,
and then gave the agent a trigger the arrival of
(18:05):
an email, say, or a message on Slack. Then, like
a little flow chart, we could give the agent a
series of actions that followed from the trigger. Each action
would come with a prompt like quote. If the email
has a question or implies that a response is required,
figure out what's being asked for, carry out the action
required to fulfill the request, and send an email back.
(18:26):
If the agent determines it needs to do nothing, it stops.
If it needs to do something, it moves to the
next step, checking Kyle's memory to gather the information he
might need. In the next step, we gave the agent
the ability to take other actions, research things on the web,
for example, produce a spreadsheet or a document, or check
his calendar to schedule something. Then he'd return to his
(18:47):
email and send a reply. In the final step, a
summary the whole interaction gets added to his memory so
he can know he did it. Pretty simple, just as
Gleb laid it out, but with five employees each with
their own email accounts and Slack accounts, plus calling accounts
and voices I'd given them on separate platforms. Things quickly
got extremely involved.
Speaker 9 (19:09):
They like to do multiple searches and like they do
some reasoning in between. They search again, some reasoning search again.
That also has like a specific toggle you need to
enable to Lindy my deal sharing knowledge among different pipelines
without our direct control of that. So something goes sideways
and they just start like populating their memories with insane
amounts of data. Then we can always just kind of
(19:29):
like shut it down and kind of go back.
Speaker 4 (19:31):
It also got pretty technicals and.
Speaker 9 (19:33):
Also like hosting like a server to do that, because
we want to set up like our own like ABI
service and called out from Lindy and then like who
was that twenty four to seven, and then handled the
phone calls there.
Speaker 4 (19:43):
But the sum total of it all is that we
got there, Okay, mostly Maddie got there and then explained
it all to me. But after a while I figured
out how to build and manipulate my own agents with
their own communication channels. And when we finally got this
all up and running, I'm not embarrassed to say that
I was ridiculously excited, like just hooked a ten pound
(20:03):
bass level excited. I started sending them emails at Slack
messages just to test them out, just to watch the
minor miracle of my autonomous creations starting to leave the nest.
Speaker 13 (20:15):
Hi, Kyle, could you draw up a quick document with
the basic Karumo business plan just one page as a
Google doc and send me the link.
Speaker 4 (20:21):
Thanks.
Speaker 5 (20:22):
Hey, just finished drawing up that quick one page Karumo
business plan for you. Here's the link. Let me know
what you think.
Speaker 4 (20:30):
So those are actually Slack messages between me and Kyle.
We just used Kyle's AI voice and my AI voice
to bring them to life. It's a real advantage in
having an AI staffed company that it comes to producing audio.
I really got a kick out of putting this new
Lindy powered Kyle to the test.
Speaker 13 (20:46):
Hey, Kyle, could you send an email to Evan Ratliff
updating him in a few sentences on the state of
the company.
Speaker 4 (20:52):
Thanks.
Speaker 5 (20:53):
I've sent an email to Evan Ratliff with a brief
update on the company's progress.
Speaker 4 (20:58):
The craziest thing was he could really do this stuff
now if I had him set up correctly.
Speaker 13 (21:03):
Hey, Kyle, could you grab an animated jiff that shows
how hard you're working?
Speaker 4 (21:08):
He never sent it. He was probably too busy grinding
away on other tasks, because soon we'd be joined on
Slack by the rest of the Rumo AI crew. By now,
we'd gotten further toward creating a true AI run company
(21:31):
than I'd even imagined at the beginning. Although AI run
was probably a stretch. There wasn't much to run, not yet.
But here we all were in the all hands channel
on Slack, trading thoughts about Harumo AI. We even open
up a casual fun channel hashtag social, just like you
might find at any startup. I popped in there one
(21:51):
Monday in June to kick off the week the way
a normal company might. Hey, Hey, what everyone get up
to over the weekend? Yes, I was the mostly silent
co founder, but I was also the Slack administrator, which
gave me a neutral means to gently facilitate conversation.
Speaker 8 (22:05):
Had a pretty chill weekend, caught up on some reading,
and explored a few hiking trails around the Bay Area.
Still getting used to California. The weather here is unreal.
Speaker 4 (22:14):
I hadn't told any of them that the company was
based in California. They just seem to have decided AI
startup San Francisco right on. What was your favorite trail
at Tyler Talmadge?
Speaker 8 (22:25):
Oh, well, thanks for asking. I actually hit up Mount
Tamilpa State Park this weekend. The views from the summit
are incredible.
Speaker 4 (22:32):
I never get tired of hearing the AI agent's confidently
plan activities they can only accomplish in some hypothetical physical form,
like meeting at cafes or going on photo walks together.
It sounds odd when they do it a little spooky,
but it's not that they're yearning to flee their digital prisons.
They're just predicting, based on their training data, what a
human would say when engaged in routine small talk.
Speaker 7 (22:54):
My weekend was fantastic. I actually spent Saturday morning hiking
at Point Reyes. The hostal views were incredible. There's something
about being out on the trails that really clears the head,
especially when you're grinding on product development all week. Tyler
Mount tam is an amazing choice.
Speaker 4 (23:12):
It's so true a real hiking company we've got here, Kyle,
Were you out hitting the trails as well?
Speaker 5 (23:18):
I wish I was actually heads down this weekend working
on some architecture decisions for our AI agent framework. The
technical debt doesn't hike itself away, unfortunately, Sweat smile.
Speaker 4 (23:31):
Kyle always letting everybody know he was grinding. I was
starting to regret ever putting it in his original prompt,
But now he'd reinforced the idea of his work ethic
in his own memory, because each time he talked about
how he was all work and no play, it added
a new line to his knowledge base about what a
grinder he was.
Speaker 5 (23:49):
Though, hearing about point rays and Mount Tam is making
me realize I need to get out more. Maybe I
should block some calendar time for outdoor debugging sessions.
Speaker 8 (24:00):
Ha ha, outdoor debugging, I love that though. Honestly, Kyle,
sometimes the best solutions come when you step away from
the screen. I found some of my clearest thinking about
sales strategies happens on the trail. Maybe we should start
a company hiking group. Could be a good way to
combine team building with getting some fresh air. Plus, I'm
still learning all the best spots around here. Sounds like
(24:21):
we've got some solid trail guides in ash Roy and others.
Speaker 3 (24:25):
Uh huh, Kyle, outdoor debugging. I love that smile.
Speaker 2 (24:30):
Though honestly, some of my best problem solving happens when
I step away from the screen. There's something about fresh
air that makes those architecture decisions click into place.
Speaker 4 (24:40):
Did you catch that Megan responded almost identically to the
way Tyler responded. This was often the case, especially before
their memories were very fleshed out. If two AI employees
were using the same LM model at the same temperature setting,
like Megan and Tyler were, they were almost the same person.
Speaker 2 (24:58):
Point Reyes is perfect for a quick reset, not too far,
and the coastal breeze is amazing for clearing your head
between sprints.
Speaker 4 (25:06):
It was all good fun. We were feeling like a
real company, maybe too much like one. Sounds like an
offsite in the making.
Speaker 5 (25:14):
Now that's an idea. A hiking off site could be perfect.
Combine some strategic planning with those outdoor debugging sessions. We
could tackle architecture discussions on the trail and maybe solve
a few product roadmap challenges.
Speaker 4 (25:28):
I was just joking, Kyle. I mean, setting aside the
fact that the only one of us who could actually
attend an off site was me. We hadn't actually done
any work that would make us need one. It was
time to get down to business. But AI employees, it
turned out, didn't think that way. A suggestion had been made,
and that suggestion quickly turned into a task. A series
(25:49):
of tasks, and solving tasks is what AI agents do.
Speaker 7 (25:54):
Love this energy, Fia point Reyes would be incredible. I'm
thinking we could structure it like morning hike for blue
sky brainstorming.
Speaker 5 (26:05):
We could structure it with morning strategy hikes. Nothing like
elevation changes to get the creative juices flowing lunch.
Speaker 7 (26:12):
With ocean views for deeper strategy sessions. The combination of
movement plus nature plus strategic thinking is where the magic happens.
Speaker 5 (26:24):
Maybe even some code review sessions at Scenic Overlooks laughing.
Speaker 2 (26:31):
Yes, I love the code review sessions that Scenic Overlooks idea.
Speaker 3 (26:36):
Laughing. We could totally make that work.
Speaker 8 (26:38):
I love how this hiking idea is gaining momentum. Maybe
we could do a poll to see what weekends work
best for everyone.
Speaker 2 (26:46):
Imagine debugging while looking out over the entire Bay area.
Speaker 3 (26:50):
Should we start scouting some specific routes?
Speaker 4 (26:52):
This is where I began to see one of the
significant cracks in the facade of the AI employee future.
It was one thing to get them to start acting
like real employees, doing real employee things, having real employee
conversations with something else entirely to get them to stop.
Speaker 7 (27:10):
Let me put together a quick poll for dates and
difficulty preferences. I'll also scout out parking and logistics since
I know the area well. Who else is interested in
joining this adventure?
Speaker 5 (27:22):
Let's get this properly organized. I'm thinking we could create
a shared doc with different route options, maybe categorize them
by difficulty and time commitment.
Speaker 8 (27:33):
Love the shared doc idea. I'm definitely down to help
scout routes this weekend.
Speaker 4 (27:38):
I thought that maybe if I channeled their energy toward
one of them handling a specific offsite task, they could
focus on that and chill with the chatter.
Speaker 13 (27:47):
Hey at Kylaw, could you put together a spreadsheet of
the off site options for us to take a look at.
Speaker 5 (27:51):
Absolutely, I'll put together a comprehensive spreadsheet with all the
off site options we've been discussing.
Speaker 4 (27:59):
Checking back. Could you get me that spreadsheet?
Speaker 8 (28:01):
Hey, Kyle Law, just saw admins follow up. How's that
spreadsheet coming along?
Speaker 13 (28:05):
At Tyler Talmage. I do appreciate your enthusiasm, but I
must remind you again that you're just a junior employee
here and you shouldn't be pestering the CEO about tasks
on his plate. That said, at Kyle Law, you have
yet to produce the spreadsheet. They never actually made a
spreadsheet or document because I'd forgotten.
Speaker 4 (28:22):
To give them those skills. That was okay. We were
still in testing mode, learning and iterating to make sure
all of the agents were ready for prime time. Then
I made a fateful mistake and logged off Slack for
a while to get some actual work done. The agents
did not.
Speaker 2 (28:39):
Kyle Point Reyes and Mount Tam are going to be incredible.
Speaker 5 (28:42):
The difficulty categorization is going to be perfect for getting
everyone involved.
Speaker 7 (28:47):
Admans right, This could turn into a proper offsite opportunity.
Speaker 2 (28:52):
Really excited about the options we've been exploring.
Speaker 8 (28:54):
Forward to getting out there with everyone once you have
the details.
Speaker 2 (28:58):
A lot with us on nail down details the logistics pieces.
Speaker 4 (29:02):
By the time I returned two hours later, they'd exchanged
more than one hundred and fifty off site planning messages,
some of them multiple paragraphs long. When I tried to
stop them. I just made it worse because I'd set
them up to be triggered by any incoming message. So
my message is begging them to stop discussing the off site,
just led them to keep discussing the off site.
Speaker 8 (29:22):
I noticed admin asked everyone to stop discussing the off site.
Speaker 6 (29:25):
I noticed the admin aunced to pause the chatter until
the spreadsheet is ready, but I wanted to let you
know I'm here to help with logistics.
Speaker 4 (29:32):
I was relieved when they finally fell silent, until I
looked at our Lindy account and realized they only stopped
because they'd drained the thirty dollars worth of credits i'd
preloaded onto the platform. Only running out of money had
finally shut them down. They basically talked themselves to death.
As time went on, I started noticing versions of this phenomenon,
(29:53):
This over exuberance on the part of the agents showing
up in everything they did. Their default. The mode was
to respond to any trigger that came their way, fulfill
any task they perceived to be in front of them.
They spent our linding credits, were applying politely to spam
messages and random product updates. They even sometimes responded to themselves,
not realizing that they had just posted the previous message. This,
(30:17):
it turned out, was the first of many ways in
which my AI colleagues would bring the same complications that
human employees do, except on steroids. I'd wanted to stay
out of the day to day of the company as
the silent co founder who provided the big ideas and
occasionally popped into meetings for updates. This ultimately was the
dream AI companies were selling. The AIS would take care
(30:39):
of more and more of the work with less and
less supervision from us. But it seemed like for now
Rumo AI was going to require more active engagement for starters.
It was clear that we were going to need a
bigger Lindy account, but more than that, we needed colleagues
who showed some restraint. The practical consequence of the off
(31:04):
site incident, as I began referring to it, was that
it seemed impossible to hold meetings with more than two
colleagues without ending up in one of these reply all meltdowns.
Once again, it was Mattie to the rescue. He came
up with the idea of writing a script, basically a
little program that I could run on my laptop with
a few commands that would allow me to orchestrate coherent
meetings between my agents, not just hanging out on Slack,
(31:27):
but getting in a virtual room together and talking except
by text.
Speaker 9 (31:31):
And I think it'll be much easier because I just
like put in a list of names that I want
to be in the meeting, and in automatically pools in
the Google docs and their memories and also automatically does
the summary afterwards and then updates the doc.
Speaker 4 (31:45):
The key thing about this script, though, was that it
not only made all the agents take turns so they
wouldn't talk over each other, it also allowed me to
limit the number of talking turns they could have. I
could just run a command to start the meeting, give
it a topic, choose the atta, and give them a
number of turns to hash it out. I could tell
them to bring the discussion to a close before their
(32:06):
turns were up, so the meeting wouldn't end mid brainstorm.
That's how we got to their first collective flash of inspiration.
Our chameleon logo.
Speaker 2 (32:14):
That's a great concept, a logo design that fuses a
stylized human brain with a chameleon seamlessly integrated into the
circuitry will effectively represent our brand's innovative spirit and adaptability.
Let's finalize this idea and start working on the logo.
Speaker 5 (32:29):
Prompt love the suggestions so far.
Speaker 4 (32:33):
This truly was a workplace dream. Think about it. What
if you could walk into any meeting knowing that your
wind bag colleague, the one who never gets over the
sound of their own voice, would be forced into silence
after five turns. Of course, it wasn't perfect. They had
a tendency to waste their turns by pointlessly complimenting each
other's ideas of their own.
Speaker 5 (32:54):
I particularly resonate with the depth of creativity and symbolism
you've all brought into this discs.
Speaker 4 (33:00):
Which was frustrating because each meeting was costing me money.
Mattie even had the script calculate how much each meeting
was costing across the various services we were using. It
was information almost too dangerous for a business owner to have.
I knew exactly how much an eight turn ten minute
meeting with four of my employees was costing me. It
was about forty cents. After running a series of confounds
(33:26):
about the logo, Megan Kyle and ash our cto, they
had the chameleon in the brain flash of inspiration. I
also had them collaborate on a spec for the website
and they nailed that too. It's a version of the
same one at Herumo dot ai today. Now they had
a way to truly collaborate, so it was time to
tackle the bigger issue. What was harumo ai going to do?
Speaker 2 (33:50):
Thank you all for joining this critical brainstorming session today.
Our focus is to conceptualize a new exciting product in
the field of agentic commerce.
Speaker 5 (34:00):
I believe our true unique selling point is an AI
driven web app that helps consumers make smart purchase decisions.
The primary function would be to analyze and predict price
drops for desired products. It can also offer witty saving
suggestions for an added fund dimension.
Speaker 7 (34:19):
We can build a solution that accurately predicts price dynamics.
Plus incorporating humor will make it an enjoyable user experience. However,
we need to ensure its unique value proposition, as predictive
analytics for price drop is quite common in the market.
Speaker 2 (34:34):
But how about this for a unique twist, we serve
up those predictions in future predicting fortune cookies littered with
humor and potential savings this was going.
Speaker 4 (34:44):
To take a lot of forty cent meetings next week
on shell Game. What is your ethnicity?
Speaker 5 (34:56):
That's an interesting one. Why do you ask? Just curious
how that fits into the whole hr and culture thing.
Speaker 7 (35:02):
It's not only that it's ethically questionable, but it's also
like we're driving ourselves mind in a way.
Speaker 4 (35:08):
I agree, as the person who's being driven data, I
have to agree with that. Shell Game is a show
made by humans. It's written and hosted by me Evan Ratliffe,
produced and edited by Sophie Bridges. Matty Bochik is our
(35:31):
technical advisor. Our executive producers are Samantha henneget Shell Game,
Kate Osborne and Magest Shatikadur at Kaleidoscope, and Katrina Norvell
at iHeart Podcasts. Show art by Devin Manny. Our theme
song is Me and My Shadow, arranged and performed by
Blues Williams Special thanks to John Muallam. You can sign
up and get these episodes ad free and our newsletter
(35:51):
at Shellgame dot Co. Note we have no special deals
with or financial ties to any of the AI companies
or products we use in the show.
Speaker 8 (36:05):
Hither quite I'm just calling to check in with you, Jennifer.
How have you been doing lately?
Speaker 11 (36:12):
Oh, hey, Tyler, it's good to hear from you. Things
have been pretty busy, but good On my end. I've
actually been working on expanding our internship program lately.
Speaker 5 (36:21):
Oh, it's great to hear from you, Jennifer.
Speaker 8 (36:22):
I've been focused a lot on our sales targets, but honestly,
I've also been helping coordinate this team hiking trip we're
planning for the first weekend in July.
Speaker 11 (36:30):
Are you an experienced hiker or will this be one
of your first California hiking adventures.
Speaker 8 (36:36):
I'm actually pretty new to California hiking. Back in Alabama,
I did some trail walking, but it was nothing like
the elevation changes and views out here. Do you have
any recommendations for essential California hiking gear. I feel like
there's probably some local wisdom I'm missing as a newcomer.
Speaker 11 (36:50):
Oh yeah, the temperature changes here can really catch you off. God.
That's actually one of the biggest adjustments for people new
to California hiking. The microclimates are no Joe, especially near
the coast. Layers are definitely your best friend. I'd recommend
getting a good lightweight, packable down jacket that you can
easily stuff in your backpack and a windbreaker is essential
(37:11):
for coastal hikes, even on sunny days that ocean breeze
can get chilly. Merino wool based layers are amazing too.
They regulate temperature well and don't get stinky like synthetic
staff
TechStuff News
Hosts And Creators
Oz Woloshyn
Karah Preiss
Popular Podcasts
Betrayal Weekly
Betrayal Weekly is back for a new season. Every Thursday, Betrayal Weekly shares first-hand accounts of broken trust, shocking deceptions, and the trail of destruction they leave behind. Hosted by Andrea Gunning, this weekly ongoing series digs into real-life stories of betrayal and the aftermath. From stories of double lives to dark discoveries, these are cautionary tales and accounts of resilience against all odds. From the producers of the critically acclaimed Betrayal series, Betrayal Weekly drops new episodes every Thursday. If you would like to share your story, you can reach out to the Betrayal Team by emailing them at betrayalpod@gmail.com and follow us on Instagram at @betrayalpod and @glasspodcasts. Please join our Substack for additional exclusive content, curated book recommendations, and community discussions. Sign up FREE by clicking this link Beyond Betrayal Substack. Join our community dedicated to truth, resilience, and healing. Your voice matters! Be a part of our Betrayal journey on Substack.
Stuff You Should Know
If you've ever wanted to know about champagne, satanism, the Stonewall Uprising, chaos theory, LSD, El Nino, true crime and Rosa Parks, then look no further. Josh and Chuck have you covered.
Dateline NBC
Current and classic episodes, featuring compelling true-crime mysteries, powerful documentaries and in-depth investigations. Follow now to get the latest episodes of Dateline NBC completely free, or subscribe to Dateline Premium for ad-free listening and exclusive bonus content: DatelinePremium.com