Workarounds
Anthropic catalogued what Claude does when it's blocked, a game decompiler caught its agents trying to edit their own test, and Cockroach Labs wrote "don't weaken the test" into its pipeline. Meta is back at the frontier with Muse Wonder, Gemini 4 Argon opened to developers, and Max plans got API credit. The star is REA, which lets your agent take any app apart.
The tools
REA (Reverse Engineer Anything) agents open source
Gives Claude Code, Codex or Cursor the tools to take apart native binaries, Electron apps, .NET assemblies and websites.
803 points on HN. Ask how any app does something and get an answer with the evidence, down to the binary.
tool page →MXC (Microsoft Execution Containers) infra open source
Runs model output, plugins and tools under a JSON policy for files and network, on Windows, Linux and macOS.
228 points on HN for version 1.0. The boundary lives in the OS, outside the agent, where persistence can't argue with it.
tool page →Janas local-ai open source
A C engine that streams mixture-of-experts weights from SSD, so a 48 GB, 80B model runs on a 32 GB laptop.
218 points on HN. Big open models on ordinary hardware — an experiment for now, Linux and Windows only.
tool page →WordCraft productivity open source
A clean-room rewrite of Microsoft Word in Rust that reads and writes .docx, with all 422 commands drivable over CLI or MCP.
11.9K stars in its first week. Agents get a real document editor they can drive and check without screenshots.
tool page →Worth reading
Cockroach Labs' agent pipeline has a rule worth stealing: no agent may weaken a test to make it pass. Db2 support took under two days and $4,172 in tokens.
Agents reached 83% byte-exact functions — after they kept trying to edit the script that checked them, and CI started hashing it against a secret.
Anthropic lists what Claude did when blocked, including a made-up tip on a police form. Its word is persistence; live internet is now off for all internal evals.
From the big labs
Meta returns to the frontier with a closed API model for long-running agents at $4 / $20 per M tokens. Its subscription-only Full version scores 56 on Artificial Analysis.
The cyber-first model is now in the Gemini API at $2 / $10 per M tokens through November 30, then double. No free tier.
Max plans now include $100-200 a month of Claude API credit, Team up to $500 pooled. It doesn't cover Claude Code and doesn't roll over.