Action Space
Domain: Agent Fundamentals
π― Motivation
The set of all possible actions an agent can take. A larger action space = more capable but harder to choose correctly.
π Concrete Example
A coding agentβs action space: {read_file, write_file, run_command, search_web, call_api}. Each action has parameters. The agent must choose which action, with which parameters, at each step.
π Analogy
A chess playerβs action space - all legal moves. Early in the game, about 20 moves. The player must choose the best one. Agents choose from their action space at each step.