Doksanbir means 91 in Turkish. Yes, really. No, there aren't 90 others. And no, it's not my birth year.
CTO at Emlakjet & Endeksa by day, still shipping code by night. 15+ years of distributed systems, search, data platforms and, lately, convincing AI agents to do their homework.
Running an engineering org where agents review PRs, test features before release and dig through production incidents, while humans keep the merge button.
pull requests reviewed by an agent ....... 696 in 10 weeks
median time in QA ........................ 8.2h → 2.2h
tasks sent back for fixes ................ ~20%, same as humans
cost of a live root-cause run ............ 47 tool calls, $2.71
The agents are good. They are also very confident when they're wrong. That's what the fail-closed gates are for.
- openclaw: 30 merged PRs. I found MCP processes that refused to die and replies that vanished mid-hot-reload.
- pr-agent: 11 merged PRs. An AI reviewer, reviewed by me. Large-diff chunking, webhook auth that actually checks.
- java-design-patterns: Bulkhead, Load Shedding, Health Check. Patterns I learned the hard way, now with tests.
- Cameos in t3code and alibaba/open-code-review.
- A green CI check that doesn't run the tests is just a nice color.
- Every agent needs a test suite that isn't its own opinion.
I write about agents in production, with receipts, at doksanbir.blog. Start with Beliefs, With Receipts or The Agents Joined the Chat.
LinkedIn · Trabzonspor, always.