Super-Agent, released by the DeepSeek Harness team on October 31, is a research prototype that watches GitHub and team chat and turns work an allowed person asks for into a pull request. Its design essay argues that a person can follow only 3 to 5 active sessions, so the bottleneck has moved to human attention, and it shows why delegation became structural: given a delegation-first prompt, a single session still delegated in only 4 of 5 runs. Reading the published commit and running its 1,197 tests shows that the host code decides who may assign work and which messages may reach a person. The worker's shell path through gh and lark-cli, the auto-merge condition and command blocking are left to prompts and shell wrappers, and the docs say they are not a security boundary. This post maps what is enforced where, then covers what to copy and what to add yourself when building a team bot on Claude Code or Codex, and how to measure the effect.

RTK says it cuts command output by 60-90% and has 85,000 stars. When UC Berkeley and Microsoft Research measured it under controlled conditions on SWE-Bench, an agent with RTK cost 9% less or 4.5% more depending on the setting, and in 12 of 30 repository × setting cells it cost more. TokenEconomizer, released by the same team, compresses no output. It cuts the costs that repeat on every turn (unused tool definitions, repository facts each session rediscovers, failure loops) and lowered cost by 18.4-30.9%. In 30 days of my own Claude Code logs, 97.9% of tokens were context read back from the cache. This post explains how an agent's bill accumulates, which line RTK, Serena, Aider's RepoMap and TokenEconomizer each reduce, how the paper's setup differs from the released plugin, and what to change first in your own setup and how to measure it.

REA (Reverse Engineer Anything) is an MCP server that connects coding agents such as Claude Code and Codex to reverse-engineering tools like Ghidra, IDA and JADX. It had 1,500 stars on October 4 and 275,000 on October 16. Reading its GitHub Discussions and Hacker News threads, the work people actually reported was mostly about software they already had: printing from a PC to a printer that only worked with a phone app, restoring old games, auditing their own services. What REA adds to earlier MCP servers is less an analysis engine than a record that attaches evidence and limitations to every conclusion. I read the repository and ran it against an Electron app I wrote and two npm packages, which surfaced what to know about output size and feature-trace accuracy. The showcases and an FPS-game decompilation project show that whether a result is right is decided by a separate check against the original. The post closes with the clean-room debate, an order for using it on your own project, and what package authors should move to the server and treat as their moat.

How do you engineer an LLM agent harness for a domain where wrong answers are expensive? Using Korea's ISMS-P (102 controls) as the vehicle, Claude Code reviews and audits through a domain-experts → interviewer → auditor pipeline. 10 reusable harness patterns + honest limits.

Public domain novel characters as AI agents. RAG + cognitive engine over the original text -- talk to Sherlock Holmes.

A native macOS meeting notes app. Real-time recording, automatic transcription, and AI-powered summaries -- all processed locally with on-device STT and LLM.