Agents in the Wild

This directory is no longer updated.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

Insight Health: 97% of critical-channel alerts closed without an engineer stepping in

Insight Health pairs two agents in its production alerts channel. One sees code and tickets. The other queries production data and masks patient information. An engineer reviews and merges each fix.

Organization
Insight Health
Industry
Healthcare software
When
Sep 2026
Evidence
Vendor case study

Reported outcome

Since Claude Tag went live, 97% of alerts in Insight Health's critical alert channel have closed without an engineer having to step in.

  • 97%
    Alerts closed without an engineer stepping in
  • three months
    Time in use

Caveats

Published by the company that sells the agent. Treat numbers as marketing until confirmed elsewhere.

  • Published by Anthropic, which sells Claude Tag, with a quote from Insight Health's CTO.
  • The page gives no alert count and no rate of wrong closes. 'Closed' does not prove each close was right.
  • The same page says an engineer reviews and merges each fix. It does not explain how that fits with the 97% figure.
  • The period is about three months. Claude Tag is in beta and runs only in channels with no patient information.

What the agent does

Insight Health builds MagicDocs, an AI referral coordinator for specialty medical practices. Its engineering team is small, and production alerts took time away from product work.

In the production alerts channel, two agents work together. Claude Tag sees the codebase and Linear, so it knows the code and the ticket history. Zeus, which Insight Health built on the Claude Agent SDK, runs in an environment covered by a Business Associate Agreement. It can query production data, and it masks patient information before the data reaches Slack.

When an alert arrives, the agents query production data, check the code and recent deploys, and compare the alert with past tickets. They report progress in the Slack thread. When they find the root cause, they open a draft PR and watch the tests. Then an engineer reviews and merges.

Why this works

  • The two agents have different access on purpose. Only one can touch production data.
  • Everyone in the channel sees both agents' reasoning.
  • Channel memory lets the agent recognize a new complaint as a known issue with a fix pending.

What to copy

  • Split access between agents so no single agent sees everything.
  • Mask sensitive data before it reaches chat.
  • Keep the merge with a person.

Starter prompt

Written by us, not by Insight Health or Anthropic. Give the agent read-only access and keep patient data out of chat.

Alert: {alert}
You can read: code, recent deploys, past tickets. You cannot read production data.

Check if a past ticket already covers this alert. If yes, link it and stop.
If no, list likely causes with the code lines and deploys that support each.
Say what production query a teammate with access should run.
Open a draft PR only. Do not merge.

Sources

  1. How healthcare organizations use Claude Tag — Anthropic, Sep 14, 2026

Added Nov 16, 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