Monitor
Watch the agent in production and improve it from what you learn.
Monitor is where a live agent reveals what testing missed. Users ask unexpected questions. They use unfamiliar wording. Their data changes after you build the agent.
Use Monitor to inspect individual executions, score live conversations, and turn reviewed corrections into learning cases. Apply what you learn in Build, such as clearer instructions, a missing skill, or a knowledge gap.
Review executions
Monitoring lists the agent's execution history. Each row represents one invocation, whether from the embedded widget, an API call, or a trigger.
Open Manage Agent → Monitoring to inspect an agent's production activity. The table includes stable identifiers for each execution and its conversation.
Execution ID
A unique run identifier. Select the copy icon to copy its full value.
Agent Name
The agent that ran.
Conversation ID
The conversation containing the execution. A conversation can include several executions.
Created At
When the execution started.
Use the agent filter above the table to inspect another agent in your tenant. This helps trace executions when one agent calls another.
Inspect an execution
Select a table row to open its execution trace. Review these details:
Skill selection — Confirm the planner selected skills that match the request.
Skill results — Find authentication errors, timeouts, and schema mismatches.
Guardrail results — See which guardrail handled the execution and why.
Step timing — Identify slow skill calls that increase agent latency.
Sample executions regularly to confirm expected behavior. When someone reports an unexpected response, use the Execution ID or Conversation ID to locate the exact trace.
Improve production behavior
Online-Eval — Score live conversations after release.
Reinforcement Learnings — Turn reviewed production feedback into reusable learning cases.
Learning cases do not enter agent memory automatically. Each one is reviewed and approved by a person first — auto-approval is intentionally disabled, so a bad or malicious input cannot quietly teach the agent the wrong thing.
Validate improvements with Evals before rolling them out more broadly.
Use account-wide governance
Monitor focuses on one agent's runs. Use Governance and Control Center to investigate activity across your account.
Observability — Review conversation volume, duration, and outcomes.
Evaluation & Quality — Monitor quality scores and review low-confidence results.
Controls & Audit — Review guardrail results, emergency controls, and account changes.
Logs & Data — Trace model calls, API requests, and database operations.
Last updated