Agentagon

Recursive self-improvement for AI agents. Improve application agents using code, traces and evaluations; remember successes, failures and production outcomes to guide subsequent attempts.

Install, open the dashboard, or connect through MCP. Detect agents, choose a goal, optionally add details, and click Go. You can also share a trace, discover issues, and choose a focused repair. A saved goal is optional for Fix. Broader goals use bounded Optimize workflows with accepted measurements and frozen comparisons.

The five concepts are dashboard, brain, memory, workflows and capabilities. The managed Codex/Claude brain supplies reasoning; validated code owns execution and evidence. Memory groups retain advisory lessons and target-agent knowledge with explicit access. Workflow behavior describes reproduction, regression gates, budgets, review and uncertainty.

Production feedback

Start with guided setup, then use the agent’s Improvements and Production views to follow changes beyond offline verification. The production guide explains opt-in local monitoring and how outcomes inform later attempts.