AI has helped Tzvetan Mikov work on more things at once. He is less enthusiastic about the context switching, joking that when ten agents are waiting for input, he looks for the one asking the easiest question.
The Hermes Team Lead at Meta joins us in the studio at Agent Conf 2026 in Warsaw for a conversation about AI coding in JavaScript engine development. With plenty of dry humor, he describes where agents help, what he checks in their work, and why their progress leaves him both excited and uneasy.
More parallel work, more context switching
Tzvetan says AI has increased both the number of tasks he works on and their scale. He describes working on five things at once, switching between them while waiting for slow models.
That creates its own frustration: he checks his phone and finds ten agents waiting for input. His way of making progress sometimes involves postponing a difficult question and finding an agent asking something easier.
Agents can help with low-level engine optimization
Tzvetan jokes that he wishes his work were too specialized for agents to understand. In his experience, they are familiar with other JavaScript engines and can help with low-level optimization.
He finds that giving a model some structure works better than letting it design something huge. Agents can also persist with repetitive optimization work, running through the night to find small improvements.
The tests pass. Then you look at the code.
"I only trust them when I don't look at the code," Tzvetan says, adding that he is only half joking. He describes the experience of getting a result that passes the tests, then inspecting the code and finding plenty of problems.
Later, he explains his review process: he reads specs carefully, alternates between Codex and Claude for review, and looks closely at tests. He also clarifies that he usually assigns one agent to one task, even while working on several tasks in parallel.
Plans for Hermes X and faster pre-releases
At the time of the interview, Tzvetan says a new Hermes release is coming soon and describes plans for Hermes X, where he intends to publish pre-releases.
He wants to make new work available without waiting for the longer release cycle through official channels. For him, that reflects how much faster he can now build features with AI.
Excitement and uncertainty about engineering work
Tzvetan enjoys being able to do more, but his feelings about AI are mixed. He describes being proud of something he made with Codex in 15 minutes, then questioning how much of his own expertise it required.
We also discuss his concerns about jobs and the future of his own role. He offers a personal prediction about where the technology is heading, while acknowledging that he can accomplish more with it now than he could a year earlier.
About the guest
Tzvetan Mikov is the Hermes Team Lead at Meta. Hermes is a JavaScript engine used by React Native. He joins this Agent Conf studio interview as a member of the React Native community to discuss how AI is changing his engineering workflow.
Resources
