# Shadowbox

- **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:** [Swarchis Kulkarni](https://cerebralvalley.ai/u/swarchis), [Bhargav Nemani](https://cerebralvalley.ai/u/bhargavnemani), [Aditya Singampalli](https://cerebralvalley.ai/u/aditya_singam)
- **GitHub:** https://github.com/swarchis/Shadowbox
- **Demo video:** https://youtu.be/PZl4mocI0WY
- **Gallery:** https://cerebralvalley.ai/e/crusoe-recursive-agents-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/crusoe-recursive-agents-hackathon/hackathon/gallery/22

Shadowbox is a system where two AI teams get better at chip design by competing with each other, while the underlying model stays exactly the same.

In the chip industry, one group of engineers designs the chip and another tries to find bugs in it, and that back-and-forth is one of the slowest and most expensive parts of building hardware. AI can already write hardware code, but it has no way of improving on its own, so we recreated that same dynamic between AI agents.

A Builder team writes Verilog circuits from plain-English descriptions, and a Breaker team writes test inputs designed to make those circuits fail. A real circuit simulator checks every result against a hidden correct answer, so scoring never relies on an AI’s judgment. After each round, a coach on each team suggests a change to its team’s instructions, sometimes writing a small Python tool to catch a specific mistake, and a gatekeeper keeps that change only if it improves the score on problems the coach has never seen. We also wrote a Verilog reference skill for the Builder, a short guide to the mistakes that most often make a design compile but behave wrongly, which the coach can switch on and the gatekeeper keeps only if it helps. On new problems without an answer key, the Builder writes five drafts and the Breaker finds inputs where they disagree, so we can pick the version most of them agree on.

Everything runs on Crusoe, using three Nemotron models along with Gemma. In our recorded run, the system went through 7 rounds in about 26 minutes with no human help, for 96 cents in total. The gatekeeper kept 3 of 42 proposed changes, and on 92 unseen problems the pass rate rose from 42% to 45%, then to 53% once the Breaker helped choose between drafts. The learned playbook also raised Gemma’s score from 81.5% to 83.7%. Most of the gain so far comes from choosing between drafts, and getting the Breaker to beat random testing is what we’re working on next.

## More from The Recursive Agents Hackathon

- [Crusoe-Cooler-System](https://cerebralvalley.ai/e/crusoe-recursive-agents-hackathon/hackathon/gallery/19)
- [Nemotron Done Means Done](https://cerebralvalley.ai/e/crusoe-recursive-agents-hackathon/hackathon/gallery/20)
- [Ming](https://cerebralvalley.ai/e/crusoe-recursive-agents-hackathon/hackathon/gallery/21)
- [Concord](https://cerebralvalley.ai/e/crusoe-recursive-agents-hackathon/hackathon/gallery/23)
- [Robot-Agent-Lab](https://cerebralvalley.ai/e/crusoe-recursive-agents-hackathon/hackathon/gallery/24)
- [Done](https://cerebralvalley.ai/e/crusoe-recursive-agents-hackathon/hackathon/gallery/25)

---

Markdown version of https://cerebralvalley.ai/e/crusoe-recursive-agents-hackathon/hackathon/gallery/22. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt