Kiro’s production-triage agent: a playbook learns a cell check and a bad lesson infects later runs

First-party operating case, August 21, 2026; checked October 23. Sai Srinivas Somarouthu, “How We Learned to Trust an AI Agent to Triage Production Incidents”, describes the team that runs Kiro’s model-serving data plane. Not a newly deployed October service and not a controlled study. This case bears on whether a saved skill transfers a lesson, a known warning left outside a planning search and discriminating controls during production debugging.

Architecture and episode. A dispatcher starts a headless Kiro CLI session per alarm ticket; the agent can use logs, tickets, Slack and scoped read-only AWS accounts via tool servers, plus steering rules, 107 on-demand skill playbooks and an archive distilled from past investigations. The operator reports 96.9% of tool calls are reads. At a Sunday 02:33 alarm with streams stalling, the agent posted a query-linked diagnosis at 02:46; a human decided the next action. In a separate incident one week earlier, an apparent model-wide issue concentrated 96.5% of errors in just two serving cells, found by joining stream failures to their routing records, and led to a per-cell quota diagnosis and an upstream dependency bug. The author says the agent proposed a rate-limit skill update including the per-cell check that night and humans review skill changes. He does not

Failure of the feedback loop. The lesson-capture pipeline stored whatever text was pending when a session was interrupted; raw ticket comments entered the corrections store and subsequent sessions treated them as guidance. The team added schema validation before storing lessons and a nightly prune; one outdated reference-doc line had also propagated to multiple tickets before human correction. Claims in a diagnosis link back to queries, but the article gives no independent precision estimate, error rate before/after pruning or fraction of diagnoses confirmed. The control is human curation of the actual retrieval path, not merely a bigger memory. The team also uses a conditional-write escalation tracker to stop parallel agents opening duplicate upstream tickets.

Authority, work and costs. The steering rules explicitly allow draft code reviews without waiting for permission, but prohibit resolving tickets or changing severity without a person. For unattended sessions, the author reports per-session task-scoped credentials with an allowlist that denies Admin, and mandatory human review of proposed patches. These are operator claims, not published access logs. A representative month yielded approximately 250 unattended investigations, median elapsed time 13.6 minutes and an agent credit cost equal to nine top-tier subscriptions. That is not an hourly human saving or customer-resolution estimate; humans read briefs, correct bad guidance, own fixes and escalation. No model dollar invoice, active human time, incident resolution latency, release follow-up or customer-retest cohort appears. This is triage → recommendation → human decision, not autonomous repair.

Comparison and test. Unlike Teotti’s lost post-commit warning, Kiro reports a deliberate correction-to-skill promotion and schema gate. Unlike Sridharan’s author-directed production control, the playbook aims to seed the next agent’s question selection; it still needs the proper incident match and a human who checks the diagnosis. Testable next artifact: link an original erroneous claim, corrected document/skill commit, new alarm and log of discovery, skill read, per-cell query, eventual repair and requester outcome. See the format specification for discovery mechanics; format portability alone neither verifies operational correctness nor restricts AWS access.