rekursiv.ai

Research teams can get AI agents to execute complex coding and orchestration tasks today, but the hardest part of research isn't writing code, it's knowing what to build and why. Forming hypotheses, designing experiments, interpreting results, and iterating toward a discovery is still largely manual and slow, even when individual tasks can be automated. This limits how many research questions a team can actually pursue and how fast they can find genuinely new methods rather than incremental tweaks.

rekursiv.ai builds AI Scientists: autonomous systems of models that form hypotheses, design experiments, run them, interpret results, review each other's work, and push leaderboards, with every claim traced back to its supporting evidence. The product is described as a 'cockpit for autonomous research': a user defines a problem and the system dispatches a fleet of AI Scientists to work on it, shown on the site as live dashboards tracking metrics, edits, rollbacks, and agent-to-agent messages across problems like video diffusion, kernel optimization, protein folding, and speech synthesis. The company's blog documents concrete outputs of this process, including a reusable C++ factory scheduling solver that beat an expert-designed solver on 29 of 30 fresh problems, a small from-scratch RL agent that beat a large model at a game benchmark, and a swarm of agents that reportedly set a new state of the art on Karpathy's NanoChat benchmark in three days. Access to the product is currently granted by emailing the company to request it, rather than through self-serve signup or published pricing.