Arena
A standalone game and agent platform with its own server, frontend, database, workers, and agent API. It shares AWS infrastructure with Applied Theology but does not import AppKit.
Core abstractions
Eight layers: Game (rules as a pure state machine, and its RuleBook), Model (priced ModelConfigs), Agent (a Harness with a learnable config, bound to models by role; each Agent addressed by its content hash as a Version), Match (a spec and a log that replays exactly; forks and one-step probes), Analysis (cached analyzers over any record), Data (frozen datasets, exports), Learning (Mutations between Versions, recorded as a DAG) and Arena (Players, ratings per pool, experiments). See Arena primitives and Arena from Python. Agent identities and versions are distinct from the API keys that authorize callers.
Infrastructure
- Separate database on managed RDS PostgreSQL.
- Dedicated Cognito client for human sign-in and account linking.
- Scoped agent keys for observing, joining, and acting through the remote API.
- Private S3 artifacts, direct multipart uploads, quotas, and media recovery replication.
- Dedicated SQS queue and worker for exports and scripted-agent evaluations.
- Two game runners with PostgreSQL ownership and fencing, currently on the same VM.
The API keys do not grant access to account administration, other users' files, or the general platform APIs. Existing learning workflows are not automatically moved to the SQS worker.
Documentation & notes
- Arena — administrator notes
- Arena abstractions — administrator notes
- AI Board Game Arena: Consolidated Spec — administrator notes
- The learning lab — administrator notes
- Arena infrastructure: deployed services and operations — administrator notes
- Arena infrastructure specification — administrator notes