Healthy
No stalls, no failure streaks and latency within bounds. The agent is doing its job.
Agent Studio
Agent Health
See every running agent at a glance. The health dashboard grades each agent on stalls, failures and latency, so you spot problems before they reach customers.
- Health signals tracked
- 5
- Grading levels
- 3
- Alert channels
- Email and SIEM
- Threshold config
- Per agent
Health grading
Three grades, one decision at a glance.
Every agent gets a grade based on recent behaviour. The grade updates continuously so the dashboard is always current.
Watch
A stall or a short failure streak detected. The agent is still running but something may need attention.
Degraded
Repeated failures or a long stall. The agent may need a profile update, a tool fix or a rollback.
What gets monitored
Five signals per agent.
The health system watches for patterns that predict problems before they become incidents.
Stalls
An invocation that stops progressing. The system detects when an agent hangs on a tool call or an approval wait.
Failure streaks
Consecutive invocations that fail. A short streak raises a watch. A long streak marks the agent degraded.
Latency
End-to-end time per invocation. A spike may mean a slow model, a slow connector or a guardrail bottleneck.
Approval wait
How long write actions sit in the approval queue. Long waits mean the queue needs attention or the agent needs fewer writes.
Eval regression
Periodic eval runs catch drift. A score drop compared to the last promotion baseline triggers a review.
Alerting
Know before your customers do.
Email alerts
Configurable thresholds per agent. When a metric crosses the line, the owner gets an email with the agent name, the metric and the current value.
SIEM integration
Alerts can forward to your SIEM in CEF format. Security and operations teams see agent health alongside the rest of your monitoring.
Improvement loop
Monitor, diagnose, improve, validate.
The Studio closes the loop. When an agent degrades, you diagnose in the invocation history, improve in the Studio and validate with evals before promoting the fix.
Monitor
The health dashboard tracks stalls, failure streaks, latency, approval wait times and eval regression across all running agents.
Diagnose
Drill into the invocation history. See every step, guardrail check, tool call and decision the agent made.
Improve
Update the profile in the Studio. Change the procedure, adjust guardrails or swap tools. Every change is a new version.
Validate
Dry run and eval the updated version. Compare scores against the previous version before promoting the change.
Integration
Cockpit and Studio together
The Cockpit gives operators a real-time control room across all agents and products. The Health dashboard in the Studio focuses on the build-test-improve loop for individual agents. They share the same data and link to each other.
See the health dashboard live.
We will walk through the health grades, drill into an invocation trace and show how the improvement loop works.