Anthropic Builds Claude Tag Agent for Internal CI/CD On-Call, Reducing Test Failure Investigation from Hour to Minutes

Title

Anthropic Builds Claude Tag Agent for Internal CI/CD On-Call, Reducing Test Failure Investigation from Hour to Minutes

Content

Anthropic's Continuous Integration team built Claude Tag, an AI agent that serves as the first responder for CI/CD test failures inside the company. Authored by CI engineer Sachin Malhotra, the agent now writes the first situation report in every recent CI incident that has one, typically publishing its first analysis within 15 minutes of an alert and a first evidence-grounded analysis a median of 14 minutes after an incident opens; in the fastest cases it names the root cause within 4 minutes. A representative case illustrates the shift. A colleague Slacked Malhotra at 10pm about roughly 44 tests on a new service that were not firing. Instead of an hour-long investigate-and-fix session, Malhotra pulled in Claude. Claude found the tests had disappeared when a feature flag was turned on that morning and that reverting the flag would be safe. Three minutes after the revert, Claude confirmed via Slack that the skip rules had been removed and the error rate was back to baseline. That round-trip replaced an hour of manual investigation. Claude Tag is the backbone of the on-call agent. It holds memory across the on-call Slack channel, provides per-turn instructions during an incident, and reacts in real time to events in that channel and others; it runs scheduled routines via natural-language prompts such as "run CI handoff every Monday at 9:00am EST." Claude Tag operates as its own service account and is wired into the tools an Anthropic CI engineer needs — Datadog, Grafana, PagerDuty, GitHub, Kubernetes, and Slack incident channels — through MCP Connectors. Standing instructions live as markdown "skills" in a GitHub repository so multiple teammates can iterate on them like code. Triage runs as a dynamic workflow: Claude Tag spins up an orchestration agent that launches executor subagents in parallel to investigate each dependency and source of truth. The orchestrator and executors are guided by an investigation skill with reference markdown files for each bug class — Malhotra cites a 617-line investigation skill for shadow divergence bugs. Executors report findings back to the orchestrator, which synthesizes a coherent SITREP in Slack. Resolution is partially autonomous through a separate Claude Code agent that handles progressive deployment behind feature flags, manages canary traffic, and ramps flags up or down; Claude Tag also opens PRs that on-call engineers review and merge. Verification closes the loop. Claude uses the same MCP connectors to verify the fix and writes a post-mortem to lessons.md; a separate "ci-weather" agent compiles incident channel data, build metrics, merge queue stats, and deploy lag into a public newsroom-style report. Anthropic notes the full setup took hours, not days, and has open-sourced a generalised on-call kit at github.com/anthropics/oncall-kit that turns a team's own incident history into triage playbooks and leaves a read-only Claude in the incident channel that diagnoses, escalates, and learns.