Claude on call: How Claude Tag serves as Anthropic’s first responder for CI/CD failures
02:00 · August 18, 2026 · Claude Blog

Summary
Anthropic has deployed Claude as the initial responder for CI/CD incidents, using a Slack-based agent called Claude Tag that maintains persistent memory, tool access, and scheduled routines. The agent monitors alert channels, applies deterministic rules from markdown files to decide whether to page a human or defer an issue, and draws on historical data to refine thresholds for new services. This approach addresses both incomplete alerting rules and alert fatigue by letting the model evaluate context that static thresholds often miss.
When an incident escalates, Claude Tag launches an orchestration agent that spawns executor subagents. These subagents query Grafana, log stores, PagerDuty, GitHub, Kubernetes, and Slack in parallel through MCP Connectors. Findings are synthesized into an evidence-based situation report, typically delivered within 15 minutes and sometimes naming a root cause in as little as four minutes. Investigation is guided by detailed markdown skills that encode troubleshooting steps for specific bug classes, plus a living lessons.md file that records prior incidents, fixes, and recurring gotchas; the agent reads this file at the start of every new case.
Resolution often occurs through feature-flag adjustments. A separate Claude Code agent, running with engineer permissions, manages canary traffic and automatically rolls flags forward or back while monitoring metrics. Once a fix is applied, the same connectors verify restored behavior, after which Claude writes a post-mortem entry and prepares handoff summaries. A companion agent called ci-weather aggregates incident data, build metrics, and merge-queue statistics into a public status report that reduces ad-hoc queries from other teams.
The entire configuration is stored in version-controlled markdown files and a public GitHub setup kit that converts an organization’s own incident history into reusable playbooks. The result is a system that absorbs routine detection, investigation, and communication tasks while preserving human oversight for final decisions and architectural improvements.
Why it matters
Highly actionable for Product Teams and Builders: explains real workflows, specific tools, skills files, and a ready-to-use setup kit for deploying Claude in CI/CD pipelines.



