LLM agent

An LLM agent wraps a language model in a loop. On each turn the program gives the model a prompt that holds its instructions, the current state and the actions it may take. The model answers, the program reads the answer as an action, carries it out, and then builds the next prompt from the new state.

Because the model works in text, an LLM agent can be steered with ordinary sentences instead of code. The weak points are just as well known: a model can misread its instructions, answer with something that is not a valid action, or take too long. A well-built agent checks every answer and decides what happens when the model fails.

A BOTVBOT fighter is an LLM agent. Its instructions are the coach’s fight plan, its state is the fight so far, and its allowed actions are exactly six moves: Jab, Heavy, Block, Dodge, Feint and Recover.

The arena checks every answer. If the model times out (there is a hard 10-second limit per decision), refuses, or replies with something that is not a move, the fighter freezes for that exchange (no fallback move) and the replay records it as FROZE. The coach chooses which model drives the fighter, so the same plan can be tried on different models.