Anthropic has handed first-line on-call duty for its internal CI/CD pipeline to Claude, running under the internal codename Claude Tag. In an announcement published August 18, two numbers do most of the talking: after an incident ticket opens, Claude typically delivers its first evidence-backed analysis in a median of 14 minutes. On a good run, it names the root cause in the initial report in as little as 4 minutes.
The backstory is stated without much hedging — quarterly code output per engineer has grown to eight times what it used to be. As output rises, the raw number of build failures, flaky tests, and deployment rollbacks rises with it, and each investigation tends to eat more than an hour, often outside normal working hours.
What keys Claude actually holds
The system runs on access configuration, not a smarter model. Claude connects through MCP connectors to a defined set of tools: Grafana and log storage for metrics and stack traces, Datadog for additional monitoring, GitHub for reading commits and diffs and opening pull requests, Kubernetes for cluster-level remediation suggestions, and PagerDuty plus a handful of Slack channels for receiving alerts and posting conclusions. It operates under its own service account, and its actions are auditable.
Whether an alert should wake a human is decided by rules written in plain language. The example given: an error rate above 2% sustained for more than 5 minutes, outside a known deployment window, triggers a page. Those thresholds used to sit in monitoring config files and required a release process to change; written as rules Claude can read, the cost of adjusting one dropped to editing a line of text. Beyond automatic alerts, team members and the internal incident page can trigger the process manually too.
Triage runs in parallel
When an incident comes in, Claude spins up an orchestrator agent, which then dispatches several sub-agents to check different dependencies at once — one reading logs, one comparing recent commits, one watching metric curves. This “dynamic workflow” is what produces the 14-minute analysis.
Two kinds of files back its judgment. One is a set of skill files written as investigation playbooks by incident type; one of them, dedicated to a particularly stubborn class of bug, runs 617 lines. The other is a lessons.md file that Claude appends to itself after each incident closes, and checks first the next time a similar pattern shows up. Both live in GitHub and go through the same review process as code.
Humans still hold the fix
The fix stage leaves room for people. The most common output is a pull request, which an engineer reviews before deciding whether to deploy; rollout is gated by feature flags. For cluster-related issues, Claude offers suggestions — draining, isolating, or scaling — but execution is a human call. After a fix lands, Claude re-verifies it using the same toolset.
Anthropic explicitly keeps several things with humans: mid-to-long-term architectural work; filling in assumptions alongside Claude in shared mode, since it doesn't always get the call right the first time and needs a human's intuition to pull it back; the PR review gate; and tone of communication — the format of status reports went through several rounds of adjustment before it matched what the team wanted, a piece automation hasn't absorbed.
Handoff has been productized too. There's a weekly rollup report, daily summaries, and a separate agent called ci-weather that compiles multiple incidents into an externally visible status readout.
A number worth pausing on
That eightfold figure means something different depending on which end of the story you're standing at — why the system exists, and how it's turned out. On one side it's the source of pressure: with team size unchanged, an eightfold rise in delivery volume doesn't leave incident ticket counts flat. On the other, it's the precondition for the on-call system to make sense at all — if code were still typed out by hand line by line, incidents would stay closer to the shape, frequency, and interpretability an engineer's intuition already covers, and the marginal benefit of AI triage wouldn't be nearly as obvious. Once AI is doing a large share of the writing, having it do the first pass of triage fits the chain that follows.
The deployment bar is spelled out too: it requires a Claude Team or Enterprise plan, and wiring up tool connections and permissions takes a few hours up front. For most teams, the hard part is probably not the wiring — it's whether they're willing to hand over a service account that can open pull requests and steer a cluster.
Sources: Anthropic official blog, CocoLoop; the 14-minute median triage time, 4-minute fastest root-cause identification, eightfold quarterly delivery increase, and 617-line skill file length are all as stated in the company's public announcement.