Maurice Heumann's agents stalled at a loose 80% until he added a pass/fail harness that compiles each function with the game's original compiler and compares bytes against the shipped EXE, then reached 99% of functions present and 83% byte-exact in about two more months. He estimates 600 to 700 billion tokens spent, and the source stays private.
Ryan Dahl says the team will fold its self hostable celld into Workers and Durable Objects, pitching them as a home for agent harnesses, while the Deno runtime gets only bug and security fixes for one more year. Deno Deploy runs for six more months with migration help to Workers for paying customers, and JSR keeps running on Cloudflare infrastructure…
Describe a game and agents build it in a live preview, with Blender integration, generated 3D assets, and an optional loop that builds and judges candidate changes, while projects stay as ordinary local files. The MIT licensed app runs on macOS and Linux (Windows pending), works through your existing Claude Code or Codex CLI or Ollama, and v0.1.6 (Oct 10) runs OpenCode 1.18 and 2.x side by side…
MOLT Sinai treats GitHub issues as patients and gives agents roles like triage nurse, fellow, and review attending, with mandatory plans, small decomposed changes, and human chiefs taking escalations. Over five months it landed more than a million lines of code on CockroachDB's migration tools, including IBM Db2 support in under two days…
The open source reverse engineering MCP server now gives agents ADB device inspection, split APK pulls, logs, screenshots, and app lifecycle control, plus JEB decompilation via a local JEB MCP server. It also adds tRPC and GraphQL call discovery in web bundles, Ghidra support for Windows DLLs and MIPS ELF, and setup for Pi, Hermes, and Qwen Code…
The MIT licensed, any model terminal coding agent now keeps one memory every project can read, syncs a navigation map at the end of each turn (view it with /map ), and shows elapsed time on MCP calls. It also makes Claude Haiku 5.5 its default Haiku tier and compacts tool results instead of truncating them…
Spotted something we missed? Start a thread.