Action (Agent)

Everything an agent does that leaves a mark outside its own text — opening a file, calling a weather API, sending an email, clicking a button in a browser — arrives as a single action. It is the point where reasoning stops being a suggestion and becomes a change.

How an action is produced and run
The model never acts directly. It emits a structured request — a tool name plus arguments that must match a declared schema — and the surrounding runtime executes it, then feeds the result back as an observation appended to the context. That alternation is the act step of the plan–act–observe loop ReAct made standard, and the unit that frameworks such as LangGraph schedule, retry and log. The actions available at any moment are the agent's action space, fixed entirely by which tools have been registered — nothing else is reachable.

Where actions go wrong
The shape of an action call is easy to get right and its content is not:

• Arguments are filled in plausibly but wrongly — a correctly formed call against the wrong record.
• A failing action is retried on every pass of the loop until the step budget or the spend cap is exhausted.
• An irreversible action — a payment, a deletion, a production deploy — executes before anyone reviews it, which is why such calls are routed through an approval gate instead.

Why it matters
A single model call can only produce words; actions are what let an agent finish work. Precisely because of that, every control an operator holds — permission scoping, sandboxing, audit logging, cost caps — attaches to the action, not to the reasoning behind it.