Agent observability: health, latency, and heartbeats that mean something

Uptime for an agent is not binary. An agent can respond quickly with the wrong tool, or slowly with the right answer. Observability has to show both liveness and usefulness.

Four signals that earn a dashboard

• Heartbeat: the process is alive and can accept work
• Latency p50 and p95: users feel the tail, not the mean
• Tool error rate: the agent’s hands, not only its words
• Containment: how often a human still has to finish the job

Heartbeats should be cheap and honest

A heartbeat that only pings the process will miss a wedged tool. A heartbeat that runs the full agent will DDoS your own bill. Check process liveness often, and run a synthetic task on a slower cadence.

Put health next to the people who ship

My Agents at six12labs is built for this: status, region, version, and latency in one view. If health lives only in a vendor console, it will not change how you ship.