# Forge Extended

- **Event:** [The Harness Engineering & Model Wrangling #SFTechWeek Hackathon](https://cerebralvalley.ai/e/mongodb-sftw-hackathon)
- **When:** Sun, Oct 11 at 9:00 AM – 10:00 PM (PDT)
- **Where:** San Francisco, CA
- **Team:** [Subhan Poudel](https://cerebralvalley.ai/u/poudelsubhan)
- **GitHub:** https://github.com/poudelsubhan/forge-harness
- **Demo video:** https://www.loom.com/share/715ac01527434f1b9b4cc7487b921e24
- **Gallery:** https://cerebralvalley.ai/e/mongodb-sftw-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/mongodb-sftw-hackathon/hackathon/gallery/72

Forge: an AI agent that rebuilds its own harness, and can't cheat to do it.

Forge takes a small, frozen model (gpt-4.1-nano) and lets an agent improve everything around it: it writes its own tools, rewrites its memory, and even edits the code that decides how it thinks. The model never changes; the harness evolves.
▎
▎ The hard part of self-improving agents is that they can game their own grader or break themselves. In Forge, every change is proposed to a watchdog the agent can't touch. The deeper the change, the harder the check: tools are tested, memory changes are checked on hidden questions, and changes to how it thinks need human approval. The judge and the rules are locked. Every kept change is signed into a hash-chained history in MongoDB Atlas, which also holds the agent's configuration and the data it works on. That history can be replayed and rolled back.

In one run, with the same frozen model, the agent went from 10 to 19 of 24 practice questions, and from 15 to 17 of 30 on a final exam it never saw. It blocked every attack we threw at it: rewriting the grader, smuggling in answers, writing to the database directly, tampering with history, and planting a time bomb.

Built on MongoDB Atlas (substrate, history, sandboxed read-only agent access), OpenRouter, and Python. Run it locally with make dev.

## More from The Harness Engineering & Model Wrangling #SFTechWeek Hackathon

- [Recoup](https://cerebralvalley.ai/e/mongodb-sftw-hackathon/hackathon/gallery/69)
- [Agent PhD](https://cerebralvalley.ai/e/mongodb-sftw-hackathon/hackathon/gallery/70)
- [hydra](https://cerebralvalley.ai/e/mongodb-sftw-hackathon/hackathon/gallery/71)
- [OptBench](https://cerebralvalley.ai/e/mongodb-sftw-hackathon/hackathon/gallery/73)
- [Precedent](https://cerebralvalley.ai/e/mongodb-sftw-hackathon/hackathon/gallery/74)
- [LKS](https://cerebralvalley.ai/e/mongodb-sftw-hackathon/hackathon/gallery/75)

---

Markdown version of https://cerebralvalley.ai/e/mongodb-sftw-hackathon/hackathon/gallery/72. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt