Manage agents (Flows)

Executions

Every run of an agent — whether triggered from the Playground, a scheduled job, an API call, or a chat session — is recorded as an execution. The execution history lets you audit what happened, diagnose failures, and understand routing behavior across real traffic.

Opening execution history

There are two ways to reach the execution history for a flow:

  • From the Flows list: click the executions icon (play circle) on the agent's card
  • From the agent graph: click View Executions in the Panel tab overview

Both open the execution list filtered to the selected agent.

Execution status

Each execution has a status that describes its outcome:

StatusMeaning
runningThe execution is currently in progress
approval_pendingThe execution is paused, waiting for a tool approval decision
completedThe execution finished successfully
failedThe execution encountered an unrecoverable error
timed_outThe execution exceeded the maximum allowed run time
canceledThe execution was stopped by a user or system action

The routing timeline

Clicking into an execution opens the routing timeline — a step-by-step trace of everything that happened during the run.

Each step in the timeline records:

FieldWhat it shows
Step numberThe iteration within the run
Event typetool_call, transfer, delegation, or route_error
Sub-agentWhich sub-agent this step belongs to
HandoffFor transfers and delegations: the source and destination sub-agents
Tool nameFor tool calls: which tool was invoked
StatusWhether the step completed, failed, or was denied
LatencyHow long the step took in milliseconds
ErrorFor failed steps: the error code and message

The routing timeline gives you a precise answer to "why did the agent respond that way" — you can trace exactly which sub-agent handled each part of the conversation, which tools were called in what order, and where any failures occurred.

Diagnosing failures

When an execution has a failed status, open the routing timeline and look for steps with a non-completed status. Common patterns:

  • A route_error step means the orchestrator produced an invalid handoff — the agent tried to transfer to a sub-agent it wasn't connected to, or the routing response couldn't be parsed
  • A tool_call step with failed status means a tool returned an error — check the error message to see whether it was a network issue, a bad input, or an application error from the target URL
  • A transfer step followed immediately by a failure usually means the receiving sub-agent encountered an error on its first iteration

Pending approvals

Executions with approval_pending status are paused, waiting for a human decision on a tool call. Open the execution to see the approval card — the tool name, inputs, and risk category are shown. Approve or deny the call to resume the execution.

Pending approvals expire after five minutes. If no decision is made, the call is automatically skipped and the execution continues with the agent being notified of the timeout.

Scheduled run history

Scheduled runs appear in the execution list alongside manual and API-triggered runs, with a schedule label identifying the source. Filter by source to see only scheduled runs, or check the schedule's own edit view for a summary of its recent fires.