Guides

These guides move from a first synchronous update to production experiment concerns. Start with the training loop, then follow the path relevant to your run:

1. Build a training loop.
2. Add telemetry and TensorBoard.
3. Overlap sampling and optimization with concurrent collection.
4. Persist the complete run with configuration and checkpoints.
5. Compare trained agents through evaluation and run concurrent searched matches with Arena.

Use the Gymnasium and PettingZoo adapters when downstream tooling expects an ecosystem environment API instead of reinfors' native Env or batched Engine.

Before tuning parallelism or search width, read the five-step collection round, then use sizing a run to choose n_games, record floors, and search budgets.