Agents in the Wild

This directory is no longer updated. Last update: November 16, 2026.

AWS's Kiro team runs about 250 unattended alarm investigations a month, with a median of 13.6 minutes

The team that runs Kiro's chat serving path lets an agent investigate production alarms with read-only tools and post a diagnosis. Only people resolve tickets or change severity.

Organization
Amazon Web Services (Kiro)
Industry
Cloud computing
When
Aug 2026
Evidence
Company reported
Agents and tools
Kiro CLI

Reported outcome

About 250 unattended investigations in a representative month, at a median of 13.6 minutes each. AWS's post gives no accuracy rate.

  • about 250
    Investigations, representative month
  • 13.6 minutes
    Median time per investigation
  • 96.9%
    Tool calls that were reads

Caveats

Stated by the company that uses the agent (blog, earnings call, interview). Not independently audited.

  • AWS sells Kiro, and this is its own product blog.
  • The post reports no accuracy or error rate. It describes trust as a process, not a measured result.
  • It describes one 'representative' month and does not say how it chose that month.
  • The post says errors got through: one stale documentation line spread into several tickets before a person caught it.

What the agent does

When a production alarm fires, a dispatcher starts a headless Kiro CLI session for the ticket. The agent investigates with scoped, read-only tools and posts a diagnosis for the on-call engineer. Each logged step links to the query behind each claim.

The agent must rule out competing explanations before it concludes. It never resolves a ticket or changes severity. Only people do that.

What happened

The team reports about 250 unattended investigations in a representative month, at a median of 13.6 minutes. In all, 96.9% of tool calls were reads.

The team also lists what went wrong. One stale line of documentation spread into multiple tickets before a person noticed. The learning loop once stored unvalidated text that later sessions treated as true, so the team added schema validation and a nightly prune.

What to copy

  • Default unattended sessions to read-only.
  • Link every claim to the query that produced it.
  • Make the agent rule out other explanations before it concludes.
  • Validate anything the agent saves for later sessions.

Starter prompt

Written by us, not by AWS. Give the agent read-only credentials.

Alarm: {alarm}
Read-only tools: {tools}

List at least three possible causes.
Run queries to test each one. For each claim, show the query and result.
Say which causes you ruled out, and why.
Post a diagnosis and a confidence level. Do not resolve the ticket or change its severity.

Sources

  1. How We Learned to Trust an AI Agent to Triage Production Incidents — Kiro (AWS), Aug 21, 2026

Added Nov 9, 2026 · Sources last checked Nov 16, 2026.

Get new use cases weekly

One short email. Sources included.

The newsletter is paused. Past issues stay in the archive.

Related use cases