I build tools for human–AI collaboration, with a focus on how people supervise agent execution and verify its results.
Main work: Ouroboros
Ouroboros is an agent OS for AI-assisted software development, connecting requirements interviews, execution, and evaluation. I contribute to its orchestration, runtime integrations, and verification layers.
My work concentrates on the boundaries where an agent's report needs to match what actually happened:
- Execution and recovery. I work on background-job polling, result recovery, and checkpoint compatibility. I unified how command-line, MCP, and iterative development runs are initialized, including state storage and worker limits. (Job lifecycle, runner assembly)
- Evidence for verification. I tighten how file-change claims are tied to execution evidence, capture tool-call records for artifact collection, and separate source checks from assertions that require an executed test or runtime observation. (File claims, tool evidence, runtime assertions)
- Runtime integration and diagnosis. I contributed the Zcode runtime and LLM adapter, and worked on installation diagnostics and MCP probes that distinguish startup, tool discovery, and cleanup failures. (Zcode, LLM adapter, installation doctor, MCP probe)
- Keeping execution aligned with intent. I added structured records of unclear requirements and their source evidence, and a check comparing the original goal with the execution summary. I also fixed model routing so a user's quality-first preference is preserved. (Ambiguity evidence, goal drift, quality-first routing)
I also explore how human decisions can be recorded, examined, and turned into action through codbeing, Pullim, and 30p / Human30.
Collaborative projects with SergioBuilds: 말결 and FlashPatch.