Proximal Labs
The large scale adoption of the Claude model family inside of Cursor was driven by the models' strong agentic coding capabilities. Recent advances in RL enable models to perform more reliably across long horizon tasks involving multiple tool calls and complex reasoning; for instance, Kimi K2's explicit focus on agentic tool calling has made the model a popular choice for developers building general purpose agents. We expect that RL-enabled improvements like these will lead to breakthroughs for many startups using the Claude API to work on agents for white collar domains outside of single-turn coding assistants.
Proximal Labs builds RL environments that broadly simulate knowledge work to best equip models with the ability to serve developers building agents of this kind. We have thought through a few selected use cases already, but are willing to explore and see what kind of environments are most needed.
Team
Justus Mattern: Studied physics at RWTH Aachen and worked on LLM robustness in the Empirical Inference group at the Max Planck Institute for Intelligent Systems. He built Revideo, an open source animation library backed by Paul Graham and General Catalyst. He also led Prime Intellect's research efforts around reinforcement learning.
Navid Pour: Studied CE at the University of Toronto and was a medalist in the Iranian Olympiad for Informatics. He was a founding engineer / researcher at Cursor where he worked on Cursor Tab, code retrieval and inline edits. He also built infrastructure for browser agents at Browserbase. Most recently he co-founded Fetchr with Calvin.
Calvin Chen: Started a logistics company that he bootstrapped to $1.5M ARR and sold for $9M. He studied business and CS at USC and worked at the Goldmen Sachs TMT banking group. Most recently he cofounded Fetchr, a startup building AI agents to automate online shopping, with Navid.
For inquiries, you can reach us at contact@proximal.rl.