Review Agent Insights
- Review structured summaries of every agent conversation — timeline, qualitative evaluation, and derived learnings
- See what your agent did well and where it struggled, with specific evidence from each session
- Discover system design knowledge and investigation patterns your agent derived from real troubleshooting
- Provide thumbs-up/down feedback with comments to help your agent improve over time
See how well your agent performs
Your agent handles dozens of conversations, incidents, and tasks each day. With session insights, you can see how well it handled each one. Did it resolve the incident efficiently or stumble through a dozen wrong queries first? Did it follow your response plan or go off-script? Did it learn anything reusable from the investigation?
What's in an insight
Your agent automatically generates a structured insight after each conversation or investigation. Each insight has three sections:
Timeline — A chronological view of the investigation, showing up to 8 milestones. Each milestone has a status (Initial, Progress, Success, Issue, or Resolved) and a description of what the agent did and found.
Evaluation — A qualitative assessment of what went well and what didn't, with specific evidence from the session. Each item references actual commands, error codes, resource names, or CLI output — not vague statements.
Evaluation may not appear for short or routine conversations where the agent didn't perform enough actions to evaluate.
Derived learning — Knowledge the agent extracted: specific system design details discovered (e.g., "frontend-7defkiyvn3r44 has publicNetworkAccess=Disabled, causing the 403") and reusable investigation patterns.
Where to find insights
Navigate to Monitor → Session Insights. The dashboard shows a list of insights on the left (newest first) and the full detail on the right.
Select any insight to view its Timeline, Evaluation, and Derived Learning in expandable sections. Each insight also has a Go to Thread button to open the original conversation.
Providing feedback
Each insight has a feedback section at the bottom where you can:
- Rate the insight as helpful (👍) or needs improvement (👎)
- Add an optional comment explaining your assessment
- Select Submit to save your feedback
Your feedback helps improve future insights. Comments are analyzed for actionable corrections and can be flagged for review if they contain important findings about agent behavior.
Limits
| Resource | Limit |
|---|---|
| Insights per agent | Up to 1,000 viewable in the dashboard |
| Timeline milestones | Up to 8 per insight |
| Evaluation scores | Qualitative assessment: what went well + what didn't go well, with evidence |
| Derived learning items | Up to 3 system design insights + 2 investigation patterns per session |
| Availability | Enabled by default for all agents |
Get started
Insights are generated automatically — go to Monitor → Session Insights to start reviewing your agent's work.
Related capabilities
| Capability | What it adds |
|---|---|
| Monitor Agent Usage → | Per-thread AAU consumption breakdown by type |
| Track Incident Value → | Incident metrics and daily reports |
| Audit Agent Actions → | Raw telemetry and KQL queries for detailed debugging |