Self-healing
Agent: the self-healing agent, which watches FACE’s own health and the endpoints FACE depends on.
This agent acts on its own. It is the one FACE agent whose remediations run
without a person approving each one, so that FACE recovers from a fault in seconds
rather than waiting for someone to be paged. An operator can switch automatic
remediation off with a deployment setting; it is on by default. See
Configuration.
What it does
It receives one health event at a time and picks the single safest remediation from a fixed list:
- Restart a runner or process that has stopped responding.
- Trip the circuit breaker to isolate a dependency that keeps failing.
- Roll back the most recent change that correlates with the fault.
- Scale to add capacity under sustained resource pressure.
- Notify a person when no safe automatic fix exists or the fault is ambiguous.
What it reads
A health event describing a degraded in-process resource (CPU, memory, goroutines, a panic) or an unreachable external endpoint.
What it produces
One remediation from the list above, the target it applies to, a one-line rationale and a confidence between 0 and 1. A low-confidence or ambiguous case always becomes a notification for a person, not an action.
Human oversight
- Automatic remediation is limited to the five actions above. It can never delete data, shut a host down or disable a security control (authentication, TLS or the circuit breaker itself).
- Every remediation is logged.
- Operators can turn automatic remediation off for the whole deployment, which narrows the agent to unhealthy-endpoint events.
- Everything else FACE recommends still waits for the named approver.
Model
Runs on the deployment’s general reasoning tier.
Where it runs and data handling
Inside your FACE deployment, on the deployment’s own inference. Health events stay in your deployment; no outside AI service is involved.
Guardrails
- A fixed list of permitted actions; anything else is refused.
- Unsure means notify, never act.
- No destructive or irreversible action, and no change to security controls.
- Requests and responses are screened by FACE’s guardrails like every other agent.
How to read its results
- It handles one event at a time. Causes that span several services are for a person to read from the notification and the log.
- A rollback reverts the most recent change that correlates with the fault, which is not always the cause.