# Modaic

- **Event:** [WeaveHacks 3: Self-Improving Agents Hackathon with Weights & Biases](https://cerebralvalley.ai/e/weave-hacks-3-self-improving-agents-hackathon-with-weights-and-biases-7014fe80)
- **When:** Jan 31 at 9:00 AM – Feb 1 at 5:00 PM (PST)
- **Where:** 400 California St, San Francisco, CA
- **Team:** [Farouk Adeleke](https://cerebralvalley.ai/u/farouk)
- **GitHub:** https://github.com/Fadeleke57/microcode-terminal-bench
- **Demo video:** https://youtu.be/FTbWSFsRAqY?si=LEb0ELkne0O1xAA9
- **Gallery:** https://cerebralvalley.ai/e/weave-hacks-3-self-improving-agents-hackathon-with-weights-and-biases-7014fe80/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/weave-hacks-3-self-improving-agents-hackathon-with-weights-and-biases-7014fe80/hackathon/gallery/44

Microcode is an RLM-powered terminal agent built with DSPy & Modaic. It features a self-reflection loop where one instance monitors another's failures during execution. Using TerminalBench, I run 89 parallel Modal sandboxes with Claude Opus 4.5. A script: reflect.py watches for failures, curates test cases/trajectories, then calls Microcode to identify failure modes and generate:  FEEDBACK.md. Microcode then rewrites its own DSPy signature based on this feedback and pushes to hub. On startup, it pulls the latest revision, retaining improvements. The self-improvement cycle: monitor failed trajectories → analyze patterns → iterate on prompts → push updates → auto-propagate. RLMs handle the long coding trajectories. Code: github.com/modaic-ai/microcode | Engine: modaic.dev/farouk1/nanocode/tree/prod

---

Markdown version of https://cerebralvalley.ai/e/weave-hacks-3-self-improving-agents-hackathon-with-weights-and-biases-7014fe80/hackathon/gallery/44. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt