Coding Agents Alpha Tracker

Two practitioner guides agree on the main point this period: agents improve when they have fast, definitive checks to run against their own work. Also covered: Simon Willison's pattern for long AI-written PR descriptions, Theo's warning about agent-built MCP auth, how Monaco structures agent work, and a counterpoint on MaxxBench.

Theo traces a reported shift in Opus 5.5's reasoning to an Anthropic summarizer rollout rather than lost capability. MaxxBench finds that models with similar DeepSWE scores split apart when tasks change slightly. Also: running Claude Code remotely through Claude Duet, a local-first T3 Code fork, and quick builds by Orosz and Karpathy.

Theo walks through how the Claude Code team gets agents to verify their work and limits automated review rounds. Also: Armin Ronacher's small agent "factory", T3 Code data showing Opus 5.5 on max reasoning fails more often than on high, and prompt and config tips.

LangChain and Google Antigravity both shipped ways to keep agents working in the background while you stay in the loop. Also covered: Cursor's build-and-deploy apps, Theo on turning agent failures into evals, the agent-found "sshhole" bug, and a benchmark-gaming warning about Kimi K3.1.

Two practitioners explain how they decide which agent-written code gets human review. Also: a wave of acquisitions of coding-agent tool makers, two new US open-weight coding models, OpenAI's agent pipeline that found and patched Linux kernel bugs, and practical tool updates.

Several practitioners describe the same loop: collect agent failures, turn them into a measurable eval, and let agents improve against it. Also: Addy Osmani's verification habits, Claude Code 2.1.317's game-rule fix, and a fake T3 Code site that ChatGPT kept recommending.

How Anthropic's Ultra plan change interacts with apps built on the Agent SDK, Theo's blind cross-model review data, LangChain's trace-to-eval improvement loop and Managed Deep Agents beta, plus OpenAI's Pando Git backend.

Armin Ronacher argues that agent-led reimplementation works best when the agent first writes tests against the original, and points to a readline port as an example. Also covered: Karpathy's HackerTube process, the return of `ultrathink` in Claude Code, and T3 Code's managed policies.

Salvatore Sanfilippo's second-chat pattern for keeping up with what your agent built, practitioners moving away from maximum effort on Opus 5.5, Arca's production harness patterns, and releases from Kody, OpenCode, T3 Code and Hex.

Anthropic removed the 5-hour limit on its top plan, and Cursor reported a tampered Claude Code extension in its marketplace. Also: a dispute over whether to compact long agent sessions or let agents retrieve from the full transcript, DHH on reading less agent-written code, and Theo on why the choice of harness matters less.