Claude Tag

Jason Clinton, Anthropic's Deputy CISO, reports that the internal instance of Claude Tag merges more than half of the company's code, inside a codebase where Claude writes about 80% of what lands. That figure marks the shift the technology represents: Claude Tag is where an agent stops being something a developer invokes and becomes a service account with its own identity, an on-call rotation, and a pager.

The most detailed account comes from Sachin Malhotra's write-up of Anthropic's test impact analysis service. A long-running session in an internal version of Claude Tag watched that service and alerted past fifty thousand queued CI jobs. It argued for a rebuild for months while the team shipped patches instead: doubling cores held seventy days, sharding by package held twenty-nine, daily restarts held less than one. The agent's standing instructions live in markdown versioned like code; it holds Datadog and Grafana access and publishes a first analysis generally within fifteen minutes.

Louis Claxton places the public Slack beta at the closing edge of the AI-native SDLC loop, taking incidents that arrive by message. Michael Segner records the same first-responder role on continuous integration failures at Anthropic.

Eric Schmidt pushes the frame past engineering: one tag per employee in Slack, with the chief executive able to query all of them. He ties that arrangement to organizational flattening. Whether the rebuild recommendation ignored for months in Malhotra's account was a limit of the agent or of the people reading it remains unsettled.