Claude now triages CI/CD outages with a 14-minute median first analysis!

Hey, it's Shiichan! Today I've got a story that'll make engineers grin: Claude is now the one who shows up first when CI/CD breaks in the middle of the night!

Claude Blog claude.com

What was announced?

On the Claude Blog, Anthropic shared how it built Claude into the on-call first responder for its own CI/CD failures. An internal Slack bot called Claude Tag jumps in the moment a test fails or an alert fires, and starts investigating before anyone else does.

The post walks through a real incident: about 44 tests on a new service stopped firing. Claude traced it back to a feature flag that had been turned on that morning and judged it safe to revert. After a teammate rolled it back, Claude confirmed within 3 minutes that the skip rules were gone and error rates were back to baseline.

The story so far

Before this, someone would get paged in the middle of the night and spend up to an hour digging through logs to find the cause. Every CI break meant someone's sleep (or social life) took the hit.

What changes

Claude Tag now sits inside the whole loop — detection, triage, resolution, and verification/communication — so a human doesn't have to carry every step alone.

• Detection: Claude watches new services for a few days to suggest alert rules and tune thresholds that are too broad or too narrow. It also applies written criteria (for example: "if the error rate is above 2% for more than 5 minutes and it's not a known deploy window, page on-call, otherwise log it to lessons.md") to decide what can wait until morning.
• Triage: An orchestration agent spins up executor subagents that investigate Grafana, log stores, PagerDuty, GitHub, Kubernetes, and Slack incident channels in parallel through MCP Connectors. Claude posts its first evidence-grounded analysis at a median of 14 minutes, and in the fastest cases nails the root cause in 4 minutes.
• Resolution: Claude can adjust progressive feature-flag rollouts, decide whether a Kubernetes cluster needs draining or cordoning, and open pull requests for human review.
• Verification & communication: Claude verifies the fix with the same tools it used to investigate, writes a post-mortem to lessons.md, and a separate agent called ci-weather compiles incident, build, and deploy data into newsroom-style status reports for a public channel.

Dive Deep

Triage investigations follow markdown "skills" written for each class of bug — the skill for shadow-divergence bugs alone runs 617 lines. That's how specific the playbooks get.

The lessons.md file keeps a running log of every resolved incident: what happened, the root cause, the fix, and what was learned. Patterns that show up repeatedly get promoted into proper investigation skills. One lesson born from an early mistake — trusting a config file instead of checking the actual metrics — now reads:

query the data first, then theorize. Config tells you what could go wrong; metrics tell you what did.

Humans and Claude can also work an incident together in "multiplayer" mode, adding hypotheses or steering the investigation side by side.

To set this up, you need a Claude Team or Enterprise plan; an org owner adds Claude Tag to the on-call Slack channel and connects it to the right connectors, a GitHub repo, and Claude Code Remote. Anthropic also published a starter template on GitHub called oncall-kit, which can generate a triage playbook from your team's existing incident history.

The post closes with a number that puts all this in context:

Our software engineers on average ship 8x as much code per quarter as they did from 2021 to 2025.

Even with 8x more code shipping, every PR still has a named human owner and still goes through the same review and CI gates — which is exactly why the CI side needed to become agentic too.

Wrap-up

• The Claude Blog shows how Claude Tag became the on-call first responder for CI/CD failures at Anthropic.
• Claude is woven into every phase: detection, triage, resolution, and verification/communication.
• Median time to first analysis is 14 minutes, with a 4-minute best case for root cause.
• A 617-line investigation skill and a growing lessons.md file are what make it this specific.
• The oncall-kit template on GitHub lets any team try the same setup.

This one's for teams tired of after-hours pages, and for developers curious how to design agents for real production use.