Claude Tag Closed an 11 p.m. Incident in 15 Minutes

Summary

Anthropic showed one night of an internal on-call shift in a 2 minute 11 second video.
Claude Tag swept metrics, deploys and feature flags and pinned the cause in about 15 minutes.
Merging and deploying need human approval, and what it can reach is decided by the channel, not the person.

Here's how our team uses Claude Tag for on-call:

The page came at 11 p.m. The alert said the error rate on the payments API had gone over 2%, and while the on-call engineer had done nothing yet, the AI agent Claude Tag was already digging into the cause. That is the opening scene of a 2 minute 11 second demo video Anthropic's official developer account posted on X on September 12. All the engineer had done was say once, a few days earlier, that whenever an alert fires in this channel the investigation should start.

What this video shows is not a boast that AI fixed an outage but a map of how far the permissions were opened and where they were cut off. Claude Tag pulled metrics, diffed recent deploys and checked feature flags, but at every point where code would be merged or deployed it stopped in front of a person. "When an alert fires in Slack, Claude pulls metrics, diffs deploys, and checks flags," the account wrote in the post, adding that it finds a likely cause and proposes a fix, which the team can approve and merge. The place in that one sentence where the subject of the verb changes hands is the design of this product.

In the 2 minute 11 second video METAL watched, the timeline runs like this. The condition that fired the alert was the error rate over the last 5 minutes exceeding 2%, and the current value on screen was 3.1%. Claude pinned the cause about 15 minutes later. Errors had stepped up from 10:58, a feature flag had been switched on just before that, and a resend worker had no cap on concurrent sends at all.

Claude came back with two fixes. One was to turn the flag off, which is safe and loses no data but leaves the feature off until someone looks at it. The other was to cap the resend worker and keep the feature on, and Claude recommended this one, suggesting a start at 16. The instruction to investigate when an alert fires was not something said on the spot; it was pinned to that channel as a standing document.

Here the product stops once. Claude was in a state where it could have opened a pull request immediately, but the engineer had set it up in advance to propose the fix only and leave the decision to a person. So it tagged the engineer and asked whether this was all right, and the engineer answered in one line: cap it and keep the feature on. Only after that answer came in did Claude cut a branch, add a test for the cap, and open the pull request.

The pull request on screen was titled as capping concurrent sends in the resend worker, and it added 4 lines, removed 1, and passed three checks. The value that changed was exactly one: no cap became 16. The engineer read that much directly, approved it and merged it, the video explains.

The code change was four lines. Then Claude also took over the place where a person would normally sit and watch the graph before going to bed to make sure it was really fixed. After the deploy it kept watching the error rate for 10 minutes before calling it fixed, and as it closed the situation out it left a note for next time: check the flags first.

The video sums it up as an hour of digging through dashboards turning into 15 minutes and one human decision. The only judgment the engineer made was which fix to take. The time spent investigating and verifying moved away, but the judgment that carries responsibility stayed on the human side.

Looked at through a lawyer's eyes, the core of this structure is that permissions are not attached to the person. Anthropic's official documentation states that "what it can reach depends on the channel you're in, not on who you are." So the same person's reach changes when they move channels, and the person who sets that reach is an Owner in the Claude organization. The documentation also states that Claude Tag starts with no access to your external systems.

The services an Owner connects during setup become an access bundle, and that bundle is attached to the workspace or to channels. Setup runs in this order: get a pairing code in Slack to pair the workspace, create an account and enter a credential for each service Claude will work in, choose which GitHub repositories it can reach, set a monthly spend limit, and launch. To use it for watching alerts, the documentation says a monitoring connection such as Datadog, Sentry or PagerDuty is needed. That is where the explanation in the video, that the tools are all their own tools connected to Claude Tag for this channel, fits in.

Under the reply Claude leaves in the channel sit a link to open the session, the name of the model it used, and a configure link, side by side. The documentation explains that clicking that configure link shows the list of external services an admin has connected for that channel, and notes that replies in org-shared channels have no such link. What it looked at to make its judgment and who set that range are placed on the same line.

Where Claude actually does the work is not the user's computer. An ephemeral sandbox is created when a conversation starts, holds the code and files, and is discarded when the conversation goes idle. And what happened in the channel stays in that conversation as a checklist that everyone in the channel can see. It means the incident record accumulates not inside one person's tool but in a place the team looks at together.

There is a line drawn through the way the money flows, too. According to the documentation, putting Claude in Slack does not add a per-seat charge, and work done in channels and conversations draws from a usage balance funded by an Owner in the organization. A spend limit caps how much of that balance can be used each billing period. Direct messages go the other way: they run on the sender's own claude.ai account, the organization spend limit does not apply, and Owners can disable direct messages organization-wide.

The seats where it can be used are narrow as well. Claude Tag is in public beta and runs on Team and Enterprise plans, on Anthropic's first-party service. It is not available on individual plans such as Free, Pro or Max, nor for third-party deployments. Work done in a channel bills to the organization rather than the individual, and anyone in that channel can hand it work right away.

Anthropic using its own product on its own team first is nothing new. METAL reported that Claude Tag, used in Slack, is coming into Claude Desktop. This video is the most concrete scene in that sequence, a shot of an on-call engineer's night actually running inside the company. What an agent that moves without approval can do is something METAL laid out in the public repository RubyGems case, and that is where the reason this product puts a person in front of merging and deploying lies.

The post drew 191,415 views, 1,710 likes, 98 reposts and 91 replies. "Every minute counts, so we love that it starts right away," the account wrote. That said, the video shows one run that went well, and it does not include what happens when Claude picks the wrong cause. That is exactly why the documentation layers channel permissions and human approval the way it does.

The value of this demo is not the number 15 minutes but the place where those 15 minutes end. Investigation and observation can be handed over wholesale, and only the moment of merging and deploying has to be left in human hands. A design that attaches permissions to the channel rather than to the person is what holds that line, and where to draw it is decided by an Owner on a settings screen. What disappeared from the on-call engineer's night is not the judgment but the hour of digging through dashboards.