Vibe Coding Discoverrrsi
View on GitHubA Python system that iteratively proposes, screens, evaluates, and selects changes to AI agent harnesses, including prompts, tools, memory, and control flow. It supports coding, document-work, and engineering-design benchmarks, with auditable Git-based candidate histories.
Use Cases
Evolve and evaluate coding agent harnessesOptimize document-work agent harnessesImprove engineering-design agentsCompare harness changes on in-distribution and held-out benchmarksTrack agent edits, measured outcomes, and evaluation costs
Built With
- Language
- Python
- Frameworks
- Anthropic Vertex AI · LiteLLM · Harbor · MCP · FastMCP
Tags
agent harnesses · self-improvement · agent optimization · benchmarking · evaluation · harness evolution · coding agents · multi-domain agents