Kasra Kyanzadeh joined OpenAI in June specifically because he wanted to be around smart people. By mid-August, he was gone. His destination: filmmaking. His reason: AI coding had made engineering lonely.
Kyanzadeh is 33. He joined Airtable when it had five employees. He later joined Watershed at a similarly early stage. After that, he founded Lightpage, an AI notebook app. He knows how to build things from scratch with a small team. That history makes his diagnosis worth listening to.
What changed in December
Kyanzadeh describes a clear before and after. Before late 2024, he was building with what he calls “tab completion models.” Those tools removed boilerplate and syntax drudgery, but engineers were still responsible for thinking through the code and the architecture.
Then, in December, something shifted. AI agents matured to the point where a single engineer could own an entire feature without much collaboration. One person. One project. The team dynamic that had defined his career evaporated.
“That removed a lot of the annoying boilerplate of writing syntax, but we were still thinking about the code and the architecture.” — Kasra Kyanzadeh on the earlier era of tab completion tools

The specific mechanics of the loneliness
This is not a vague complaint about vibes. Kyanzadeh described three concrete problems with how AI-agent-driven work actually feels day to day.
No shared stress, no shared wins
He has a theory about how teams bond: stress. When every engineer is running their own solo project with their own agents, the shared pressure that creates real relationships disappears.
“Human bonding happens best through stress. When every person is working on their own project, you don’t get that same feeling of: We accomplished this thing together.”
The flow state problem
Each agent task takes a few minutes. Kyanzadeh called this “the worst amount of time.” Too short to do something meaningful. Long enough that you want to kick off another task. So you spin up five or six parallel threads. Or you open Twitter. Neither is flow.
Code review has become the job
Reviewing AI-generated code and checking for errors has grown as a percentage of total time. Kyanzadeh’s read on this is blunt: most engineers never liked code review. It was always overhead. Now it’s a bigger slice of every day.
OpenAI wasn’t the problem
Kyanzadeh is careful to separate his frustration with AI-agent-driven development from any criticism of OpenAI itself. He joined to work on the Codex desktop app. He described OpenAI’s employees positively and said the company had the usual social infrastructure, including happy hours. But none of that substitutes for the daily texture of collaborative work.
“To me, OpenAI was an experiment of: What is it like to build software on a large scale in this new era?” he said. “It turned out, I don’t feel like I’m doing my best work anymore.”
He finished the project he was working on before he left. The team was kind about it.
He’s not alone
Kyanzadeh’s experience matches a broader pattern. Anthropic engineering leader Fiona Fung has said AI agents made the company’s programmers lonelier. Other developers have told reporters that AI tools created workplace paralysis. The efficiency gains are real, but the human cost is showing up in early departures and career pivots.
The transferable lesson
If you’re running a small team and leaning hard into AI agents for speed, the productivity math probably looks good. The social math deserves equal attention. Kyanzadeh’s story suggests that the collaboration and shared ownership that come from working on the same problem together are not soft benefits. They’re what keeps engineers engaged long enough to do their best work.
He wanted to be around smart people. The agents didn’t give him that.

