ClickHouse says GroeneAI, its autonomous engineering agent, has contributed more than 3,000 merged pull requests to the database project since beginning work in March. The company describes an agent that finds bugs, investigates them, writes and tests fixes, and follows its pull requests through review, with people deciding whether to accept the changes.
ClickHouse Watch analysis
What happened
In a post published on 29 October, ClickHouse says GroeneAI works around the clock on failures from continuous integration (CI), fuzzing and user reports. It can also be assigned a bug by a maintainer. The agent investigates a problem, builds ClickHouse, runs tests, opens a pull request, monitors CI and responds to review comments.
ClickHouse says GroeneAI ranked second among contributors by authored non-merge commits over the three months to 19 October, behind co-founder and CTO Alexey Milovidov. The company also describes its setup: a persistent task system stores the investigation and its next steps across sessions, while separate model sessions review the plan and the resulting code.
One example in the post is a June bug in ClickHouse’s query-condition cache. The agent reproduced a failure involving pending data mutations, added a fix and tests, then addressed reviewer feedback. ClickHouse says the change took 16 days from failed test to merge. The post says GroeneAI revised its pull request five times over 12 hours after initial review findings.
Read ClickHouse’s account of GroeneAI.
Why it matters
This is a concrete account of an AI agent doing more than generating a patch: it carries work through investigation, testing, review and follow-up. The reported pull-request total suggests the approach is being used at considerable scale inside a major software project.
The less glamorous details may be the more important ones. ClickHouse describes checkpoints that preserve what has been tested and what remains to do, and a separate review process that checks the code rather than simply accepting the worker’s account of its own work. That is a useful glimpse of the machinery needed to make long-running automated coding work reviewable.
Our read
Three thousand merged pull requests is an eye-catching tally, but a merge count cannot tell us how much human effort each change saved or how many required substantial correction. The more persuasive evidence is the process ClickHouse describes: reproduce the bug, show that the test fails without the fix, and put a maintainer between the agent and the merge button. That is autonomous work with a review trail, not a robot being handed the keys and wished luck.
What to watch
- How many of GroeneAI’s pull requests are accepted, revised or rejected.
- Whether ClickHouse publishes more detail on reviewer workload and the cost of operating the system.
- Whether the approach transfers to projects with different codebases and maintainer practices.
Discussion spark: Would you trust an AI agent to take a bug from report to pull request if a maintainer still reviews every change, or should some parts of that workflow remain human-led?
Sources and evidence
- Source update (29 October 2026, 12:53 UTC)
not affiliated with or endorsed by ClickHouse