# The last harness

- **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:** [Alazar Shenkute](https://cerebralvalley.ai/u/a8nova)
- **GitHub:** https://github.com/a8nova/the-last-harness
- **Demo video:** https://www.loom.com/share/50caffc5c6e64c9daef25e15688f89cb
- **Gallery:** https://cerebralvalley.ai/e/crusoe-recursive-agents-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/crusoe-recursive-agents-hackathon/hackathon/gallery/87

A self-improving agent harness for hardware-specific tasks, where real devices decide what improves. A team of Nemotron agents ports layers from new Hugging Face models (Nemotron-H) to C, and optimizes llama.cpp CPU kernels, for four real devices: Apple M2 Pro, NVIDIA Jetson Orin Nano, Galaxy S20 (Snapdragon 865) and a Crusoe AMD EPYC CPU instance. The kernels are the ones that set LLM decode speed on these chips: the Mamba2 state update, the conv1d decode step, the MoE router, and the quantized and fp16 dot products llama.cpp runs for every token.

The learning signal is what the hardware measures, never an LLM's opinion. Every kernel is checked for correctness on hidden inputs, including edge-case sizes. On the device itself the verifier records cycles per call against the plain-C baseline, instructions per cycle, CPU time, energy where the device reports it, and which SIMD width the code actually used. It also tracks how the agents got there: wasted hardware tries, time to the first correct version, and model cost.

After each round, a reflector rewrites the team's own harness: prompts, playbook, tool-output wording, model settings and team structure. A change is kept only if the devices measure the team as at least as correct and faster on held-out kernels. Every lesson goes into a shared memory with two layers: facts true for one chip, and facts verified across chips that every agent reads. Because the signal is just a number the hardware reports, the same loop can optimize for prefill throughput, peak memory footprint or energy per token by changing what the verifier measures.

## More from The Recursive Agents Hackathon

- [CASCADE](https://cerebralvalley.ai/e/crusoe-recursive-agents-hackathon/hackathon/gallery/84)
- [Zen](https://cerebralvalley.ai/e/crusoe-recursive-agents-hackathon/hackathon/gallery/85)
- [Capy](https://cerebralvalley.ai/e/crusoe-recursive-agents-hackathon/hackathon/gallery/86)
- [FindingNEMO](https://cerebralvalley.ai/e/crusoe-recursive-agents-hackathon/hackathon/gallery/88)
- [Alchemist](https://cerebralvalley.ai/e/crusoe-recursive-agents-hackathon/hackathon/gallery/89)
- [Tornado](https://cerebralvalley.ai/e/crusoe-recursive-agents-hackathon/hackathon/gallery/90)

---

Markdown version of https://cerebralvalley.ai/e/crusoe-recursive-agents-hackathon/hackathon/gallery/87. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt