# Darwin

- **Event:** [The Recursive Agents Hackathon](https://cerebralvalley.ai/e/crusoe-recursive-agents-hackathon)
- **When:** Sat, Oct 17 at 9:00 AM – 10:00 PM (PDT)
- **Where:** San Francisco, CA
- **Team:** [Sunny Rodrigues](https://cerebralvalley.ai/u/sunnyrodrigues), [Nithin Chintala](https://cerebralvalley.ai/u/nchint)
- **GitHub:** https://github.com/srodrift/darWIN
- **Demo video:** https://www.loom.com/share/e667607b6c6e4cc498f782888ab12136
- **Gallery:** https://cerebralvalley.ai/e/crusoe-recursive-agents-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/crusoe-recursive-agents-hackathon/hackathon/gallery/37

Darwin Inc. — Human org charts are generation zero.

The problem: Every company is about to deploy teams of AI agents, but they organize them by copying human org charts — managers, reviewers, approval chains. Nobody tests whether that structure actually works for AI. Bad structure wastes money, adds loops, and ships wrong answers.

What we built: Darwin turns an organization into code. Each org is a YAML file describing teams, roles, who reviews whom, and who decides (tests, a leader, a council, or a router). We start with five human-style designs: a council of peers, a hospital-style team, a research lab, a skunkworks, and a flat startup. Each org runs real coding tasks with NVIDIA Nemotron on Crusoe, and every answer is graded by hidden tests.

The recursive part: A strategist agent (Nemotron Super) reads why each org failed and rewrites its structure, such as adding a reviewer, splitting a team, changing who decides, or spinning up an isolated skunkworks. A strict gate keeps a change only if it scores higher, or the same for less money. Survivors become the next generation. Finally, a held-out exam on brand-new tasks checks whether the evolved org really generalizes.

Results: In our full live run, 12 structural changes were tried and 1 survived. On unseen tasks the evolved org scored 0.77 vs 0.75 for the best human design and 0.70 for a solo AI worker. Evolution beat the best human design in 2 of 3 live runs; in the third, the gate correctly kept the human design. The margins are small, so we treat this as early evidence. The full run cost about $0.11.

How you see it: Darwin HQ, a 3D view where each generation is a floor, gold marks the winner, and you can replay any org working a task. Story mode explains it in plain words; Explain mode shows the YAML, the diff, and the gate math. "Ask Darwin" answers questions about the run and suggests new test cases, grounded in the real data.

Who uses it: Any team running AI coding agents. Describe your setup as one YAML file, bring tasks with tests (or your repo history), run evolve with a hard spend cap, and get back the best structure, the diff, and the evidence.

Built with: NVIDIA Nemotron 3.5 Lightning (workers) and Nemotron 3 Super (strategist) on Crusoe Managed Inference; You.com search is wired in per role. 260 tests, open source: github.com/srodrift/darWIN

## More from The Recursive Agents Hackathon

- [THE agents](https://cerebralvalley.ai/e/crusoe-recursive-agents-hackathon/hackathon/gallery/34)
- [HomeOS](https://cerebralvalley.ai/e/crusoe-recursive-agents-hackathon/hackathon/gallery/35)
- [Proofline](https://cerebralvalley.ai/e/crusoe-recursive-agents-hackathon/hackathon/gallery/36)
- [Harveyj](https://cerebralvalley.ai/e/crusoe-recursive-agents-hackathon/hackathon/gallery/38)
- [Escape Labs](https://cerebralvalley.ai/e/crusoe-recursive-agents-hackathon/hackathon/gallery/39)
- [Team Sway](https://cerebralvalley.ai/e/crusoe-recursive-agents-hackathon/hackathon/gallery/40)

---

Markdown version of https://cerebralvalley.ai/e/crusoe-recursive-agents-hackathon/hackathon/gallery/37. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt