Agenta
Lead Product Engineer · open-source LLM engineering platform, now an agent builder · Berlin
I owned the frontend’s technical direction and led the frontend team: the architecture, the library choices, the shared package layer.
- I rebuilt the prompt playground in my first two months, the interaction design as well as the code. It had stuttered on one prompt with a few test cases, and came out running several prompts across many test cases side by side. Prompt management, evaluations and testsets followed.
- When the product pivoted to an agent builder, I built the surfaces it runs on: composing an agent from skills, MCP tools and a model, a streaming chat to drive it, approval gates on consequential actions, and scheduled runs.
- I designed how a running agent asks its user for things mid-run: typed forms with accept and decline instead of free text, which cuts input errors on the user's side and gives the agent structured interactions without planning every one in advance. The forms speak the MCP elicitation dialect, so connected tools can request typed input too.
- I carved ten shared workspace packages out of the monolithic app so desktop and a new mobile client can render one definition, delivered as a numbered stack of small reviewable PRs. Most of that layer is merged, the rest in review.