Kayba

Kayba is a self-improving agent framework that analyzes execution traces from your AI agents to detect failures, surface actionable insights, and apply recursive improvements. It integrates with popular coding agents like Claude Code and Codex, uploading traces and generating structured insights across categories like policy gaps, missed steps, and hallucinations. Measured on τ2-bench, Kayba has demonstrated up to 100% improvement in agent consistency over multiple iterations. The core framework is MIT-licensed and open source, with a hosted Pro dashboard available for teams.

• Recursive Reflector — automatically analyzes agent traces to detect failure patterns and generate improvement insights without manual review
• Skillbook Generation — builds a structured knowledge base of learned behaviors and policies from past agent runs
• Failure Detection — spots wrong parameters, skipped policies, bad routing, and hallucinations before they reach end users
• Insight Dashboard — surfaces insight categories, severity levels, and frequency distributions across your agent's trace history
• Coding Agent Integration — call Kayba directly from Claude Code, Codex, or any coding agent via CLI to upload traces and fetch improvements
• LiteLLM Integration — supports multiple LLM providers through LiteLLM, making it model-agnostic
• LangChain & Browser-Use Support — integrates with LangChain pipelines and Browser-Use agents for broad framework compatibility
• Async Learning — supports asynchronous trace ingestion and improvement cycles so agents can learn without blocking production
• Pipeline Engine — a built-in pipeline engine with branching, parallelism, and custom step support for complex agent workflows
• Team Collaboration — Pro plan includes a hosted dashboard with team collaboration features and bring-your-own-API-key support