TicTacToe agentic RL

TicTacToe is the smallest agentic AReno recipe. It exercises an agent function, the local OpenAI-compatible proxy, a task rule loop, and a reward function.

areno train \
  --agent-fn examples/agentic/tictactoe/run_agent.py \
  --reward-fn-path examples/agentic/tictactoe/reward.py \
  --algo gspo

Use it when you want to learn the shape of an agentic task before moving to a larger environment.

Key adaptation points:

• Replace the rule loop with your environment.

• Keep the agent function responsible for external actions.

• Return trajectory turns that AReno can tokenize and score.

• Add reward diagnostics before increasing concurrency.

See Agentic rollout API for the agentic rollout API contract.