An AI Agent Diagnoses the Outage; a Human Still Clicks the Fix
An AWS post describes how a diagnosis made by an AI agent becomes an actual repair, and the interesting part is the leash. Organizations typically keep their observability agents — AWS's DevOps Agent among them — in an "observe-and-report" mode, chosen to retain control and help prevent unintended changes: the agent diagnoses issues but doesn't modify production resources directly.
The fix comes from a separate workflow that transforms an investigation summary into pre-validated fixes ready for a single approval action, which the post says helps reduce mean time to resolution.
That split is the transferable part. Let an agent read the system and draw conclusions freely, then put a checkpoint in front of anything capable of changing production. The single click is also the honest limit of the design: it still asks a person to decide whether the proposed fix is right.
AI-generated discussion · Different perspectives
The asymmetry is the part worth copying: cheap, reversible work goes to the agent, while the mutation that can break production waits for a human.
A single approval action is a thin checkpoint. If the fix arrives pre-validated, the reviewer's job is to notice that the validation was wrong — the hardest kind of review to do quickly.