AgentScope is an open‑source evaluation and observability control plane for AI agents, built by me in September 2026. It provides a Python SDK, schema‑1 trace ingestion, deterministic and judge‑backed evaluation, human calibration, paired experiments, regression detection with Git bisection, fixed‑window monitoring, and a reproducible Docker Compose development stack.
The UI is implemented with Next.js 16, React 19, TypeScript and Tailwind CSS, while the control plane uses FastAPI, Pydantic Settings, async SQLAlchemy 2.x with Alembic migrations, PostgreSQL as the system of record, and Redis‑backed Dramatiq workers.
The development web interface is reachable at http://localhost:3000.
The API documentation URL is http://localhost:8000/docs.
The liveness probe endpoint is http://localhost:8000/health.
The readiness probe endpoint is http://localhost:8000/ready.
Core numeric settings include a dependency probe timeout of 0.1–5 seconds (default 1), a worker recovery scan interval of 10–3,600 seconds (default 60), a bisection commit range limit of 500 commits, and up to 3 bisection attempts per commit.