The creator of Claude Code handed most of his apps’ maintenance to an agent: 388 PRs opened, 180 already in production

🤖 The creator of Claude Code handed most of his apps’ maintenance over to Claude.

In a few weeks: 388 pull requests opened, 180 already shipped to production.

A Slack channel where the agent shows up every day

The story comes from Boris Cherny, the creator of Claude Code.

He put Claude in charge of a dedicated Slack channel, where every day it runs automated routines across all his apps: iOS, Android, desktop, web and CLI.

The four routines that run every day

We’re talking mostly about the kind of ongoing maintenance work that tends to pile up:

• Crash fuzzer: it opens the app in a simulator, uses it at random until it crashes, finds the root cause and proposes a fix.
• Duplicate unifier: it looks for near-identical abstractions in the codebase and opens a PR to consolidate them.
• Dead-code removal: it deletes unreachable code. When it’s unsure, it first adds a log to check whether that code really runs.
• Fixing problematic abstractions: it spots patterns that are starting to cause trouble and proposes changes.

How much of that work actually reaches production

The numbers after a few weeks:

• 388 PRs opened automatically
• 180 PRs already approved and shipped to production
• review with Claude Code Review + human oversight
• roughly 1 PR in 50 is noise to throw away

The numbers come from Boris Cherny’s post of 13 August 2026, where he describes the proj-claude-maintains-apps Slack channel and the daily routines across iOS, Android, Desktop, web, CLI and Agent SDK. The Decoder picked up the figures the next day: 388 pull requests opened and 180 merged works out to a 46% merge rate.

And when that happens, Boris fixes the routine that generated the PR. That way the same mistake doesn’t come back.

Another interesting detail: most of the routines run on Opus, while the more powerful models are saved for the more complex tasks.

The method, which is the repeatable part

💡 The most interesting part, to me, is the method.

How can we get an agent to look after an app without launching it by hand every time?

1) Turn repetitive work into routines

Crash fuzzing, hunting for dead code, updating dependencies and recurring checks are all great candidates.

We can run them with Claude Code’s /loop or, even better, with a cron job on a VPS.

I use this approach today to automatically monitor my own apps and my clients’ apps.

A daily routine builds up value over time.

Maintenance routines are part of how I run the custom software I build for companies: the code stays under watch after release too.

2) Keep a review step before the merge

Automated checks plus human review, especially when data, payments or security are involved.

The agent prepares the change. The merge stays under control.

3) Fix the routine when it gets it wrong

If a PR is useless, it pays to understand why it was generated and update the routine.

Every mistake can turn into a new rule for the system.

4) Use different models depending on the task

Routine maintenance rarely needs the most powerful model available.

The expensive one can be saved for the hard cases.

When maintenance runs while you’re not working

The interesting shift comes when maintenance keeps running even while we’re not working.

The next day we find PRs already prepared, ready to be checked and approved.

Are you still doing the repetitive work on your code by hand, or do you already have a routine or loop running on its own?