MCP 2026-07-28 is live

Anthropic just gave a peek behind its own curtain, and the number they shared is the kind that makes other engineering leads nervous. The Claude Code team says Claude now writes 65% of the code produced by its product team, including large chunks of the tool that runs Claude Code itself. That's not a demo stat or a cherry-picked benchmark. That's their actual, everyday shipping pipeline.

The most interesting part isn't the percentage, though. It's how deeply Claude has been wired into the parts of engineering that used to eat entire afternoons: incident response and bug triage. When a page fires, someone just tags Claude in the incident thread. It pulls the relevant graphs, diffs the last deploy against the one before it, and comes back with a root-cause theory and the name of whoever wrote the offending code. A human still approves the fix in-thread, but from there Claude opens the pull request, lands it, watches the metric climb back to normal, and resolves the page itself.

Bug triage works the same way, minus the 3am adrenaline. Claude sits in the team's feedback channel, picks up reports as they land, traces them to the actual code path, reproduces the bug, runs git blame to find who touched it last, writes a fix, and tags the responsible engineer. By the time a human looks at it, the only job left is reviewing and merging.

What's notable here is how narrow the human role has become in both loops. Approve the plan, review the patch, done. Anthropic is essentially showing its own dogfooding numbers as proof that an agent can own the unglamorous, high-friction parts of software maintenance — the stuff senior engineers historically hated doing anyway. Whether that 65% figure generalizes outside a company whose entire product is Claude is the real open question, but as an internal case study, it's a pointed one.