Claude Tag

Jason Clinton, Deputy CISO at Anthropic, reported on 21 July 2026 that the internal instance of Claude Tag merges more than half of Anthropic's code, in a codebase where Claude writes roughly 80% of what gets merged.

The operating description is narrower than that figure suggests. Tag runs as first responder for CI/CD on-call: a dedicated service account, read access to Datadog and Grafana, and standing instructions written in markdown and versioned like code. Its first analysis is generally published within 15 minutes. Michael Segner's 20 August 2026 guide to Claude Code for startups files this under a single rule, automating the tedious, with the engineers keeping the judgment cases.

The Slack framing matters more than the plumbing. Tag is in public beta there, with Claude present in the channel under its own identity, able to watch a channel or source and start tasks proactively rather than waiting for an invocation. Incident response begins in a conversation. Louis Claxton's 21 August 2026 playbook puts Tag at the closing of the Maintain loop, extending to incidents that arrive by messaging the same pattern a deterministic threshold breach already triggers. Claxton advances no measured results.

Boris Cherny lists Tag on 16 July 2026 among the multi-agent interfaces (Agent view, Desktop, mobile) that mark the step where an engineer stops supervising one agent and starts orchestrating many. What is described publicly is an access boundary and a set of versioned instructions, not a capability claim.