Running an Agent

Execute agents manually, monitor progress, handle HITL approvals, and manage budget limits

Overview

Once an agent is created and configured with tools, you can execute it manually or as part of a symphony/playbook. During execution, the agent processes its task, invokes tools, and may request human approval at defined gates.

Manual Execution

Select the Agent

Navigate to Agents and click on the agent you want to run.

Provide Input

Enter the task description or input data in the Run Input field. This is the prompt that tells the agent what to do.

Click "Run"

Click the Run button to start execution. The agent begins processing immediately.

Monitor Progress

The run dashboard shows real-time progress including:

  • Current step and total steps taken
  • Tool invocations and their results
  • Elapsed time vs. time limit
  • Token usage and cost estimates

Human-in-the-Loop (HITL) Approvals

HITL approval gates allow you to review and approve agent actions before they execute. This is critical for high-stakes operations like database writes or external API calls.

Configure HITL Gates

In the agent settings, enable HITL for specific tools or actions. When the agent attempts these actions, execution pauses and waits for approval.

Review Pending Approvals

When an agent hits a HITL gate, you receive a notification. Go to the agent's run view to see the pending action details.

Approve or Reject

Review the proposed action and click Approve to continue execution or Reject to skip the action. You can also provide modified parameters before approving.

Budget Limits

Limit Type Description Behavior When Reached
Time Limit Maximum wall-clock time for the run Agent execution stops gracefully, partial results are saved
Step Limit Maximum number of tool invocations Agent cannot invoke additional tools, must produce final output

Event Logs

Every agent run produces a detailed event log. The log records each decision the agent makes, every tool invocation with input/output, HITL interactions, and the final output. Use the event log for debugging, auditing, and improving agent prompts.

Tip: Enable HITL approval gates during initial testing of any agent that writes to external systems. Once you are confident in the agent's behavior, you can selectively remove gates for trusted operations.
Important: Agents continue to consume LLM tokens during execution. Set appropriate budget limits to prevent unexpected costs, especially for agents with web search or database tools that may trigger many iterations.

Next Steps