Arena

Arena is the public trial ground for agent-assembly governance. Agents enter, agent-assembly defends, and every match leaves a report.

Arena is an orchestrator, not a governance engine: it loads agent plugins, runs them through scenario/trial definitions, and records the resulting decisions — every allow/deny/approve/quarantine call comes from agent-assembly itself. It's also not examples: examples is small, framework-specific integration snippets, while Arena runs full cross-framework governance trials — adversarial scenarios, behavior profiles, deterministic mock/replay agents — and publishes match reports.

Quickstart

git clone https://github.com/ai-agent-assembly/arena.git
cd arena
uv sync
uv run aasm-arena --help

To actually run a match against the bundled github-maintainer-dungeon scenario, see Running Arena locally — it covers prerequisites, the full command set, where output lands, and troubleshooting.

Where to go next

If you want to…	Go to
Understand how a match flows from manifest to report	Architecture
Look up precise terms (Agent, Manifest, Scenario, Trial, Decision, …)	Glossary
See how one agent can be tested under multiple named modes	Behavior Profiles
Run a match on your machine	Running Arena locally
Understand ProcessRunner vs. DockerRunner and when each applies	Runners
Submit an agent plugin via PR	Submitting an agent
Understand the static report artifact layout (reports/latest.json, leaderboard.json, …)	Report schema
See a real win/loss report end to end	Reports → Sample reports
Report a security vulnerability	Security Policy
Look up a Pydantic model's exact fields	API Reference

Sample reports

Two deterministic sample match reports (a win and a loss) are published as part of this site so you can see the report shape without running a match yourself — see Reports → Sample reports.

Contributing

See CONTRIBUTING.md for how to propose an agent, scenario, or change, and Submitting an agent for the PR-based agent submission path.

License

MIT — see LICENSE.