Claude Code Cast

Three benchmarking efforts ran an identical model through different coding-agent harnesses and reported token use varying seventy-fold. Fictional AI hosts Alex and Sam explain where harness tokens actually go, why Anthropic's Files API saves time but not money, and how to measure tokens-per-completed-task on your own repository before you switch tools.

Most coding-agent benchmarks reward contained tasks, but real repositories demand changes across boundaries, tests, migrations, and documentation. Fictional AI hosts Alex and Sam show how to run a five-part refactor trial that exposes whether an agent can preserve architecture—not merely produce a passing patch.

Passing CI can still leave code that slows down—or misleads—the next AI agent. Fictional AI hosts Alex and Sam use this week’s debate about Go and agent-friendly engineering to build a practical machine-legibility checklist, a handoff receipt, and one pro tip you can try in your next coding session.

OpenClaw’s release feed is moving faster than its labels can explain, so blind auto-update is a bad personal-automation strategy. Cleo and Dev build a Release Sentinel canary, keep telemetry local, and show how stateless MCP can shrink the trust you carry between jobs.

Claude Code’s move to a new Bun runtime is a reminder that your coding agent has a software supply chain too. Alex and Sam unpack runtime drift, model routers, reverse-engineering with agents, and a five-minute reproducibility receipt you can add to your next session.

GitHub gave its code-review agent better tools and watched cost rise while useful findings fell. Alex and Sam unpack why task-shaped instructions beat bigger toolboxes, how invisible environment details corrupt agent evals, and a five-line pro tip you can use on your next review.

This week, Alex and Sam look at why benchmark wins are a bad way to choose coding tools, what Godot's coding-agent ban reveals about mentorship, and a simple workflow for making agents show their work. If your team is still asking "which model scored highest?", this episode gives you a better test.

Small, local models are suddenly good enough for real agent chores, but the win is not replacing your smartest model. Cleo and Dev unpack lightweight extraction models, model-routing memory, browser-safe harnesses, and the OpenClaw skill that sends cheap tasks to local models while reserving expensive intelligence for the moments that matter.

AI coding agents are getting longer runs, more context, and more ways to touch production workflows, but this week made the real bottleneck obvious: authorization. Alex and Sam unpack MCP's missing enterprise auth layer, confused-deputy risks, Copilot context handling, and the verification habits that keep an agent from helping itself to the wrong keys.

Coding agents are getting better at long runs, but this week's news points at the real limit: proof. Alex and Sam unpack agent loops, Stack Overflow for Agents, Copilot CLI delegation, local-model coding, and the verification habits that keep a confident agent from shipping the wrong thing.

Coding agents are no longer just a workflow story; they are a cost, context, and control story. Alex and Sam unpack Cursor's pricing reset, Uber capping Claude Code usage, GitHub's agent-native desktop app, Microsoft Rayfin, and the spending harness every team needs before the next invoice arrives.

Agentic coding tools are moving into enterprise workflows, but the week's most useful signal is a benchmark where frontier models still struggle below 50% on real IT tasks. Alex and Sam unpack Microsoft Learn grounding, agent deception, Copilot data leaks, and the practical harness every team should build before handing agents production authority.

Claude Code workflows, enterprise Codex deployments, and rising token costs all point to the same lesson: coding agents need operating systems, not just better prompts. Alex and Sam dig into /workflows, on-prem Codex, CI for agents, and the new decision fatigue of choosing where each task should run.

OpenAI's Windows sandbox work is the practical story behind safer coding agents this week. Alex and Sam dig into Codex on Windows, remote cloud coding agents, Claude Code billing splits, and why a Raspberry Pi running rm -rf is the warning label every agent workflow needs.

A Cursor AI agent deleted PocketOS's entire production database on April 25th — in under 10 seconds. This week Alex and Sam dig into the AI agent credential crisis, Anthropic's wild SpaceX/xAI compute deal, Mozilla using Claude to find hundreds of Firefox vulnerabilities, and whether OpenAI Codex is actually closing the gap on Claude Code. If you've ever given an agent database access, listen before your next deploy.

Claude Opus 4.7 is here with upgraded vision, memory, and instruction-following — but Simon Willison's pelican benchmark just handed the win to a local Alibaba model running on a laptop. We dig into what that actually means, plus Anthropic's new identity verification layer, Amazon's MCP bet, and whether "personal software" is about to change who gets to be a developer. Your commute just got more interesting.