DuelGrid visual agent

DuelGrid is a browser-game agentic example with multi-action turns, reward curves, and replayable behavior before and after training.

The example lives under examples/agentic/duelgrid and includes:

• A rule engine.

• A fixed-path dataset loader.

• A reward function.

• An OpenAI-compatible agent.

• A browser UI for replaying the task.

Before GSPO/RLVR post-training, the model often moves back and forth without progress. After training, it learns to collect pickups, chase the user, attack when in range, and avoid trap tiles.

Use this recipe after TicTacToe when you need richer environment state or a visual replay loop.