Three months ago, a new "coworker" joined our GitHub organization. It never complains, works through the night, and occasionally goes home mid-task when its Spot instance gets reclaimed.
Last week I ran the numbers, and they were wilder than I expected.
64% of our merged PRs were written by AI
Between April 8 and July 6, 2026, our org's ten main repositories merged 3,781 pull requests. Of those, 2,424 (64%) came from branches created by AI agents.
The breakdown:
| Type | PRs | What it does |
|---|---|---|
| Task-execution agents | 859 | Reads a Notion ticket, implements, tests, opens a PR, gets it merged — end to end |
| Maintenance bots | 1,413 | Doc sync and periodic housekeeping, quietly |
| Auto-fix agents | 108 | Watches Sentry, sends a fix PR when production breaks |
| Other agents | 44 | Odd jobs |
Humans (effectively 2–3 of us) do reviews and direction. The hands on the keyboard are mostly AI.
Median time from "PR opened" to "merged": 2 minutes
For the 859 task-agent PRs:
- Median: 2 minutes
- p75: 9 minutes
- 93% merged within an hour
"Who reviews a PR in 2 minutes?" — another AI does. When a PR opens, an AI code reviewer files findings with severity levels, CI runs, and on approval the PR auto-merges. Humans watch the results scroll by in Slack.
To be honest about the metric: this clock starts after the agent has already implemented and passed local verification. Still, from "ticket marked Ready" to "merged," you can brew a coffee and it's done before you're back.
Of course, it breaks in hilarious ways
So this doesn't read like a suspicious success story, here are real incidents:
-
Sudden resignation: agents run on Spot instances. When one is reclaimed mid-task, all that's left is a note reading
failed at step ''. Reason: (empty) - Fake Done: the agent reports "Completed!" — there is no PR. There is also no one to interrogate
- AI traffic jam: an AI-authored PR waits for an AI review to be merged by an AI, and another AI merges into the same file first
- Time-bomb tests: an agent hardcoded a date into a test; 90 days later it detonated and blocked every PR in the repo
Each of these became its own postmortem article. The incident pipeline doubles as a content pipeline.
The machinery
We run this on Codens, our own family of AI dev-automation products:
- Green turns a conversation into a structured PRD
- Purple picks up tickets and runs agents: implement → verify → PR
- Orange auto-reviews every PR (with security audit)
- CI + approval → auto-merge
- Production error? Red detects it and sends a fix PR
- Blue handles QA and E2E
Yes, this whole article is a dogfooding report for our own product. But every number above is measurable via the GitHub API.
What's left for humans
After three months, the human jobs are:
- Deciding what to build — ticket quality is everything; vague tickets get vaguely implemented
- Designing verification — "tests pass" is not enough; agents will happily cut corners unless you assert file existence and forbidden patterns
- First-response when things break — and the incidents become blog posts
It feels less like "AI taking our jobs" and more like "suddenly managing a direct report who files 2,400 PRs a quarter." Management is hard.
Codens has a 14-day free trial (no credit card): codens.ai
Even adopting just Red turns "3 AM Sentry alert" into "a fix PR waiting for you at breakfast."
Top comments (0)