Autonomous Maintenance
An AI started
fixing its own apps.
Claude Code has begun taking over day-to-day maintenance of Anthropic's own apps. Of the pull requests it generated without human authorship, 46% actually got merged — the first time a coding agent's real-world track record has been shown in hard production numbers rather than benchmark scores.
From word of mouth
to a hard number
Until six months ago, a coding agent's real ability could only be described by benchmarks or word of mouth. This is the first time raw production numbers have been made public.
According to Claude Code's creator Boris Cherny, Claude has spent the past few weeks running daily maintenance routines across Anthropic's entire app lineup — iOS, Android, desktop, web, CLI, and the Agent SDK — coordinated through a dedicated Slack channel called "proj-claude-maintains-apps," as reported by The Decoder. The work is the kind that quietly eats developer time: bug fixes, refactors, and dependency updates.
Over those few weeks, Claude generated 388 pull requests. After automated checks and human review, 180 were merged — a 46% rate. Cherny has also told Fortune that on some days he oversees tens of thousands of AI agents at once, and this 46% figure is best read as a report card from that day-to-day operation.
Why this number
matters right now
Unlike a SWE-Bench score, this measures whether a human actually approved the work in production.
A benchmark score measures whether a problem can be solved at all. 46% measures something different: whether the work passed human review and actually shipped to production. Anthropic has separately reported that over 80% of new production code was Claude-authored as of May 2026, and this 46% figure extends that story by putting a number on reviewer-side quality for the first time. This wasn't a one-off demo — it's the result of the same routine repeated every single day, which is what makes it heavier than a benchmark headline.
Who benefits, and how
For engineering teams, this is a reference point for deciding what to hand off: dependency bumps and minor refactors — the "someone has to do it but nobody wants to" work. Cherny says Claude usually gets it right on the first try, and when it doesn't, the team tweaks the routine's prompt and lets it try again the next day — an explicit improvement loop.
Business and PM leads now have an actual number to put in an adoption memo, instead of a vague sense that "agents seem to help." That said, 46% is Anthropic's own internal figure — a different codebase or industry may not reproduce it, so it should be read with that caveat in mind.
It wasn't built to take
the work away —
it was built to shrink it.
What to do next
Start with low-risk, repetitive work
Hand off dependency updates and minor refactors first — tasks where a miss costs little. That's exactly what this case covered: bug fixes, refactors, and dependency updates.
Keep human review in the loop
46% is the merge rate after review, not an unreviewed auto-ship rate. Keep the two-stage setup — automated checks plus a human reviewer — intact.
Log rejection reasons and feed them back next day
Anthropic's team reportedly adjusts the next day's routine based on why a PR didn't get merged. Treat it as an iterative loop, not a one-shot judgment.
Not an unqualified win
First, 388 is still a small sample, so the 46% figure could easily shift as more data comes in. Second, this is Anthropic's own internal measurement of its own product — it hasn't been independently verified by a third party.
Around the same period, Anthropic itself flagged concern over "recursive self-improvement" — the pace at which Claude writes code accelerating faster than expected — and raised the need for an option to pause frontier development. So while 46% is a genuine milestone, how far to extend an agent's autonomy remains a decision that still sits with humans, and that context is worth reading alongside the headline number.