Evaluate and improve real LLM applications and AI agents inside Claude Code or Codex. Define success with the user, run evaluations, diagnose failures, change prompts/tools/code, and verify improvements against frozen criteria. Also supports AI tracing and monitoring.
Evaluate failures, implement authorized application fixes, and verify unchanged acceptance criteria.
Review existing evaluation evidence and generate an offline report without calling the target.
Evaluate an existing AI application against its acceptance criteria without fixing application behavior.
Start or resume guided Evals intake to learn the target, current setup, acceptance rules and desired mode.
Instrument or inspect actual model, retrieval, tool and agent traces for the requested application.