Skip to main content
A run records one execution of an agent, task, workflow, or evaluation. Its ID connects the input, current state, events, output, and errors, independently of the client connection.

​
Definitions and executions

An agent definition contains instructions and capabilities. A task or workflow definition describes implemented work. Starting a run executes a selected definition with particular input and a runtime target. For example, the same support agent can answer many requests. Editing its definition and inspecting an earlier run are separate operations. Schedules and webhooks can start work automatically, with a new run for each execution.

​
Progress and outcome

A newly accepted run is not yet a successful result. Its recorded status distinguishes work that has not started, active work, waiting, and terminal outcomes. Status describes the lifecycle. Output and errors explain its result. Events add the history of text output, tool activity, and state changes; they can be replayed after a client reconnects. See the AG-UI event reference for streamed agent events.

​
Waiting and resuming

An agent run can wait for a tool result, human input, or another supported resume condition. A support agent that requests missing information can remain waiting while the person responds. The waiting state identifies what the execution needs. Submitting an ordinary message does not necessarily satisfy a structured input request. The applicable response or resume operation determines how execution continues. Agent, task, and workflow runs can resume a confirmed manual pause. Project-owned workflow runs also accept approval and event signals for a matching wait. Agent runs accept tool results and integration-connected signals. Other run kinds and waiting states do not all expose the same lifecycle operations. A cancellation request addresses a particular run. Inspect its recorded state and any child runs to understand the resulting execution tree. A closed stream or disconnected client is not a cancellation signal by itself. The Execution API exposes shared run operations. Evaluation-specific definitions and reports belong to the Evaluations API.