AYA
Built at The Recursive Agents Hackathon · Oct 17, 2026 · San Francisco, CA

AYA turns a feature request and a working software reference into a tested component for your own product. You provide a target GitHub repository, the behavior you need, and a reference implementation or binary. Hermes investigates it with REA inside OpenShell, builds the component, and repairs failures. Independent tests and CI check the result before it becomes a pull request. Our demo adapts a Fireworks SQL workflow for Crusoe. We also demonstrated binary investigation by recovering a JSON export feature from jq. The second part improves how the work gets done. We integrated Sakana AI’s MASS workflow search, recorded agent activity in Langfuse, and used verified answers for a real Crusoe training job. The checkpoint is trained; deployment and independent evaluation are the next steps. The longer-term goal is to reduce the cost of finding features customers will actually use and pay for. Each attempt should leave the organization with reusable tests, working approaches, failed experiments, and the reasons behind its decisions. Customer feedback would then guide both the next feature and the workflow used to build it.