Researched and written with 312,338 tokens (in 10 · out 33,316 · cache-read 65,545 · cache-create 213,467) across 5 Claude calls

Claude Code and AI code review

• Claude writes ~90% of Claude Code's code: merged PRs more than doubled June to September while the revert share fell over 40%, which the team credits to a rebuilt review process.
• How the Claude Code team does code review: the post optimizes for the least human review time that doesn't raise reverts or incidents. Commenters doubt the 30k-merged-changes-per-day figure and warn that letting engineers close security false positives without AppSec sign-off could add vulnerabilities.
• Review process thread: a second agent reviews the diff with fresh context before the PR opens. On every PR, parallel agents find bugs and separate agents vote on whether each finding is real.
• Addy Osmani, Morgan Westlee Lunt and Qing Wang, quoted: "did a human read the code?" is a weak proxy for "what could go wrong, and which check is right to catch it". Review moves from an end-of-road gate to something that runs throughout.
• Code Review now GA for Team and Enterprise: it's faster and catches more of the bugs that matter, and admins can set Quick, Standard or Deep review levels per repo.
• Claude Code v2.1.302: adds CLAUDE_CODE_AUTO_MODE_MODEL to pick the client-side auto mode classifier model, and an Idle compaction toggle in /config.
• Claude Code gzips each turn's conversation: responses start sooner because the payload reaches the model faster, and cloud sessions now spin up as soon as you start typing.

Agent harnesses, multiplexers and plugins

• T3 Code out of alpha: the ADE/multiplexer is now a stable release.
• T3 Code "pretty good now": Theo's quick endorsement of the same release.
• ADE and multiplexer tier list v2: a retest on real tasks that puts T3 Code 0.1.0 on top, with Orchestrator V2 stable and support for Pi and any ACP provider.
• Parallax: a single place to run every coding agent you use.
• Tetriz Harness: turns what a team learns into a shared agent setup for everyone.
• cursor/plugins: Cursor's official plugin spec and first-party plugins, each a standalone directory with a .cursor-plugin/plugin.json manifest.
• android/skills: Google's modular Android best-practice instructions as SKILL.md files, so agents get grounded on platform patterns.
• Plugin for subagents on other Claude accounts: it switches accounts automatically when one hits its limit.
• jev inside Claude Code's harness: an attempt to cut token maxing in Claude Code.

Agent workflow patterns

• Agents building state machines: Matt Pocock says a state machine or reducer in agent output makes the code more robust and later changes easier to test, especially on the frontend.
• GRILL.md for /grill-me: a proposed file to customize planning questions (security, module strategy) and to choose one-at-a-time or batched questioning. It would also replace /grill-with-docs.
• 13 internal tools built from specs: Claude passed every check but the build exposed 126 gaps in the written specs, so the specs were the weak link.

CLI and terminal tooling

• CLI-Anything: generates agent-native CLIs for existing software, with a CLI-Hub (pip install cli-anything-hub) for community-built ones.
• Toad: lets you point an agent at any machine you own.
• bun lint: coming in the next Bun, an ESLint-compatible linter said to be 26x faster, with over 1,000 built-in rules, that reads eslint and oxlint configs.

LLM mechanics

• KV cache explainer: the model reuses stored attention keys and values for existing context and computes them only for the new token, instead of recomputing the whole context each step.
• NVIDIA Nemotron Labs Dory 1.0: a 4B model on Hugging Face that loops 1, 2 or 4 times per token with the same weights, so you tune compute without retraining or swapping models.