Using a "chaos loop" to harden the Claude Docs launch · Vivek Nair, Anthropic · claude.dev, November 17, 2026
Anthropic ran a two-week Claude loop against staging before the Claude Docs launch
From September 4, about two weeks before launch, Nair had a Claude Code session drive a staging cluster at over a million synthetic docs a day. The loop picked a failure scenario, ran the harness, read telemetry through the Datadog MCP, proposed a fix, shipped it to staging and repeated. A coordinator session handed work to sub-agents, usually about 11 at a time in the busiest week and 30 at the peak. Before the run, Nair told Claude which decisions needed his approval: "basically anything hard to undo", such as schema or key-layout changes.
The team also picked its stack for the agent's benefit. It wrote the service in Rust and vendored its few dependencies: "We wanted any number Claude measured to trace back to code we wrote." In one case, a large doc was using about 3 GB on its pod; profiling traced it to every keystroke copying the doc into several data structures. After a copy-on-write fix the doc topped out around 180 MB, and the pod estimate for launch fell from more than 500 to about 50.
Anthropic reports that users created over 4.6 million docs by September 22, with 60 server errors across 2 billion requests, 629 production rolls in September and no rollback. Nair puts the inference cost at roughly $125,000 at Claude Fable 5.2 list price. He estimates a similar push on a small stack at $3,000 to $10,000, and says he has not run it. The post does not include the underlying telemetry, and we have not seen the figures reported independently.
Why it matters: This is a worked recipe for an agent-run load test with a stated price: give the agent telemetry, write down the invariants, and reserve hard-to-undo changes for a person. The argument that runtime choice should follow what the agent can measure is a new criterion for picking a stack, and the post says the team did not compare Rust against Go directly.