On August 22, Anthropic engineer Sachin Malhotra publicly described an internal system called "Claude Tag" — a cyber on-call agent — in a report covered by Dazhiyuan (新智元). The system can localize failures in as fast as 4 minutes and produce an incident report in 14 minutes, and more than 80% of its connected codebase was written by Claude itself. The full methodology has been open-sourced as the oncall-kit project.
How Claude Tag Works
Claude Tag is not a single-point demo. It embeds Claude permanently into the "nerve endings" of enterprise collaboration:
- Runs as a persistent session inside Slack, connected to alert sources such as Grafana, Datadog, and PagerDuty.
- Reads alert meanings in natural language, then performs cross-source correlation analysis to identify the specific service, specific commit, and responsible owner.
- Pushes a situation report with remediation suggestions to the on-call engineer.
- It works very well on known failure patterns (CPU saturation, memory leaks, dependency timeouts).
- For novel, unknown failures — an unprecedented kernel panic or cloud-provider-level jitter — it still depends on a human making the call. Anthropic's own engineers stressed that "Claude is a very capable copilot, but does not vouch for reliability."
The whole pipeline compresses a traditional on-call workflow — typically two hours or more — down to minutes.
80% Self-Produced Code
The more notable fact is not speed but that over 80% of the code merged into oncall-kit's main branch was generated by Claude itself, with the engineering team only reviewing and signing off. Anthropic isn't just using Claude for new features — it has pushed Claude into the operations loop, letting it modify its own deployment stack. This pairs with the Claude Code 2.1.234 release on August 21 (browser-use + Skills API + Files API), forming a closed loop: "can write code" on the front end, "can do on-call" on the back end.
Why On-Call Is the New Moat
On-call moves AI coding from a "personal productivity tool" to an "enterprise infrastructure component." For two years, AI coding companies competed on generating more lines of code at the PR stage — the "writing" track. But code spends a long operational life in production; for a mid-sized internet company, operations costs are typically 3–5x development costs. Anthropic's move signals: "Our model can not only write code, it can take on operational responsibility — and it wrote the operations tooling itself."
Meanwhile, Codex surpassed 20 million weekly active users (confirmed by Tibo on X, August 22). The Claude Code vs. Codex rivalry, once framed as "who writes complex code better," has shifted: while Anthropic pushes Claude into on-call, Codex is still competing on user volume and data-retention pledges. The battleground has moved from "model capability" to "deployment stack capability."
Limits and Competitive Pressure
Don't over-romanticize. oncall-kit's 4-minute localization has boundaries:
The Deeper Play: A Data Feedback Loop
oncall-kit creates a data feedback loop. Traditional AI coding vendors lack real production incident samples — what breaks in production, how it's fixed, whether rollbacks are clean — all locked inside enterprise monitoring systems. By trading open source for community-contributed incident samples, every company adopting oncall-kit becomes one of Anthropic's "training grounds." This is an early sketch of shifting the business model from "selling model tokens" toward "selling an operations operating system."
Bottom Line
The real AI coding news of August 22 is not Codex's user count, but Anthropic putting Claude on on-call duty. In the same week, Google Antigravity Anywhere shipped remote agent access (Aug 21) and OpenAI pledged not to retain enterprise data — all three major players moved the battlefield from the model layer to the deployment layer. This week's marginal shifts are more worth recording than any single new model release.