aec-bench Documentation
aec-bench is a Python platform for creating, running, evaluating, and improving Architecture, Engineering and Construction benchmark tasks for AI agents.
Use these docs to move from a first local run to maintaining reproducible benchmark datasets, agent harnesses, evaluation traces, and leaderboard-ready results.
Recommended paths
- New to aec-bench: start with Quickstart.
- Writing artefact tasks: read Tasks, Templates, Contributing Tasks, and Datasets.
- Writing staged or stateful tasks: read Finite Lifecycles and Interactive Worlds.
- Running agents: read Harnesses, Configuration, Prime Agent, and Environment.
- Reviewing results: read Scoring, Traces, Classification, and Review and Reporting.
- Improving agent workspaces: read Evolution, Agentic Variation, and Swarm.
- Developing harnesses and governed workflows: read Meta-Harness Composition and Adaptive Harnesses.
- Publishing or integrating: read Library Catalogue and Prime Lab.