Action Space
The full set of moves an agent is allowed to make at any given step.
Think of it like
The legal moves in a board game — you can only do what the rules permit, however clever your plan.
Example
A browsing agent’s action space is click, type, scroll, and navigate; it can’t do anything outside that menu, no matter what it wants.
How it actually works
The action space defines the agent’s reach — which tools it can call, which operations it can perform. Its size is a real design tension: too small and the agent can’t solve the task; too large and it gets lost choosing, or gains powers you didn’t mean to grant. Constraining the action space is also a core safety lever, since an agent literally cannot do what isn’t in it.
For product teams
Bound this deliberately — it’s both what the agent can achieve and what it can wreck.
For engineers
The enumerated set of tools and operations available to the agent at each decision step.
Related
- Tool Calling — A single choice drawn from the action space.
- Environment — The world the actions operate on.
- Approval Gate — Fencing risky actions behind approval.
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome