Skip to main content

Architecture Diagrams

High-level diagrams explaining the key architectural innovations in the Agent Evals framework.

Diagrams​

1. System Architecture​

Hexagonal architecture showing layer organization and dependency relationships. Core defines port interfaces, adapters implement them, with strict dependency inversion enforced.

2. Evaluation Modes​

Unified interface for both live evaluation (execute agents) and historical evaluation (score saved traces). Same framework, scorers, and result format for both modes.

3. Adapter Architecture​

Platform and scorer adapters are fully independent and extensible - any platform works with any scorer combination without vendor lock-in. Third parties can add custom adapters without modifying core code.

4. CI/CD Integration​

Automated quality gates in CI/CD pipelines catch agent regressions before production, preventing "whack-a-mole" problems in development.

5. Evaluation Lifecycle​

End-to-end workflow from identifying failure modes through building eval packages, execution, and iterative improvement. Covers both live and historical evaluation modes.