Rajan Agarwal
I work on coding models at OpenAI, on leave from my undergrad at Waterloo. Recently I've worked with great people and teams at Thinking Machines, Proximal, Amazon AGI, Thoughtful Lab, and Cohere Labs.
Reach out! I always want to meet more people with controversial opinions, fun stories and new ideas, or who want to play tennis. I am most active on:
Notable Work
- FrontierSWE(Apr 2026)
Ultra long-horizon coding agent benchmark for engineering and research.
- Amazon Nova Act(Dec 2025)
RL for Amazon Nova Act, a browser-use agent. Co-led enterprise performance.
Related:Amazon RL Training - LLMs invent compression(Nov 2025)
Discrete compression schemes learned by LLMs for long-context compaction.
Related:Deep Learning with Yacine - Natural deception with RL(Nov 2025)
Hidden-information game agents trained with RL and studied for emergent deception.
Multilingual alignment through encoder injection under Cohere Labs.
- Shadow(Aug 2025)
Open-source background coding agent that runs in isolated environments. 1.5k+ Stars.