Theodor Marcu

Theodor Marcu is Head of Product Growth at Cognition, the company behind the Devin coding agent. He has described a shift from handing an agent a one-off migration to making delegation part of everyday engineering. When several agents can work at once, engineers need a way to see which jobs are still running and which need their attention. That problem helped shape Windsurf’s command center for local and cloud agents.

From redistricting maps to consumer AI

Theodor grew up in Romania and studied computer science at Princeton. In 2019, he and four classmates started Representable as a course project. He worked on engineering and product for the open-source redistricting tool. Residents can draw the boundaries of their community and describe the interests they share. The resulting maps and descriptions give district mapmakers something they can use when deciding whether a proposed boundary would split a community.

Theodor worked as a software engineer at Retool before starting and later selling Superdrop. There, he and John Suh built Cami, a multilingual AI assistant that people could message on WhatsApp. Theodor reports that it grew to more than a million monthly active users. Superdrop also built DreamTavern, a platform for AI storytelling and roleplay.

Superdrop’s apps used separate libraries to log their interactions with models, duplicating code. Theodor drew on that experience to build PromptFactory, an open-source library that records context alongside prompts and replies. Its design lets developers add support for new model providers without writing a separate logging library for each app.

Where coding agents cost engineers time

• Waiting for an agent. Theodor co-authored Cognition’s SWE-1.5 release write-up, which explains how faster model inference exposed other delays in Windsurf. When an agent spends less time generating code, lint checks and command execution can account for more of the wait. Cognition reworked those parts of Windsurf alongside the model speedup.
• Keeping track of several agents. Cognition engineers were already planning tasks locally and handing them off to cloud agents. Theodor described how switching among those sessions took enough attention to warrant a different interface. Windsurf 2.0 puts local and cloud sessions on one board so engineers can see what is running and where they need to step in. Theodor also proposes tracking sessions through to merged pull requests so teams can judge how much agent activity turns into code they actually merge.