The Health-Check Pattern: Probing Agents Before They Work
A fleet should not discover its agents are broken while they are working. The health-check pattern is the discipline of probing agents before, during, and after their work — a lightweight verification that the agent is alive, configured, and capable of the task it has been given. Health checks turn “trust the agent” into “verify the agent,” and they are the operational form of the zero-trust discipline (S7.3).
The three probes
Pre-flight: before an agent starts a task, probe its readiness — is the profile loaded, are the credentials valid (S7.5), is the toolset reachable, is the model endpoint alive? The pre-flight is the health check before the heavy lift (S12.7), and it fails fast: a broken agent is detected in seconds instead of after an hour of wasted work. In-flight: during long tasks, the progress heartbeat (OP-3) confirms the agent is advancing, not spinning. Post-flight: after the task, verify the artifact — the post exists, the image imported, the log is structured (OP-7).
The probe as a small task
A health check is itself a small task — which means it needs the same discipline as any task: a bounded cost, a defined timeout, and a clear failure mode. The probe should be cheaper than the work it protects: a readiness check that costs more than the task is a tax, not a check. The health-check pattern is also how the fleet scales: new agents are probed before they join the pool, and degraded agents are probed out of rotation before they poison a batch.
Health as a fleet signal
Health checks generate the fleet’s vital signs: uptime per profile, failure rate per tool, readiness per task type. The signs feed the dashboard (S6.7), the cost meters (OP-4), and the incident response loop (OP-5). A fleet with health checks knows its own state at every moment — the same self-knowledge the simulation layer provides for business models (SIM-3). Health checks are how the fleet stays honest about what it can actually do.
Grounded in the OP operations series, the S7.3 zero-trust article, the S12.7 health-checks article, the OP-3 watchdog article, and the S6.7 dashboard article. Eighth article in the Round D operations track.
