rekursiv.ai
rekursiv.ai provides an autonomous research platform that operates as a team of AI Scientists—specialized agents that hypothesize, design experiments, run them, and interpret results without human intervention. The system has demonstrated independent discovery of novel algorithms, achieving 100% accuracy on Sudoku and 71.4–75.5% on ARC-AGI at dramatically lower computational cost than comparable LLM-based approaches. Users define a problem through a cockpit interface and monitor live experiment fleets with per-claim evidence tracing.
- Artificial Intelligence
- AI Agents
- Software Only
Funding
Raised to date
$500KRaised to date based on public sources. This may differ from the amount the company actually raised and is based only on what is publicly available on the internet.
Founders
Founder details are not available yet.
Product
Problem
Modern AI systems can execute tasks and write code, but they struggle to discover new knowledge. They rarely form novel hypotheses, design sophisticated experiments, or interpret unexpected results—the core activities of the scientific method that drive major breakthroughs. This limits their ability to solve open research problems autonomously.
Solution
rekursiv.ai provides an autonomous research platform that operates as a team of AI Scientists. Users define a problem, and the system dispatches multiple specialized agents—including Scientists, Analysts, Engineers, and Reviewers—who collaboratively hypothesize, design experiments, run them, and interpret results. The system can run fully unattended or accept human guidance. It continuously pushes leaderboards while tracing every claim back to its evidence. The AI team has demonstrated independent discovery of novel algorithms, achieving results like 100% accuracy on Sudoku and 71.4–75.5% on ARC-AGI at dramatically lower computational cost than comparable LLM-based approaches. Each experiment cycle runs through a message-routing script that enables the agents to pivot when evidence contradicts assumptions, including abandoning its own hypotheses after finding quantitative counter-evidence in literature.
Target Audience
Primary users are researchers and research organizations working on machine learning, constraint satisfaction, and general reasoning problems who need to scale their experimentation capacity or explore novel hypotheses beyond human throughput.
Features
- Multi-agent architecture with distinct roles (Scientist, Analyst, Engineer, Reviewer) that coordinate via a message-routing script, enabling autonomous hypothesis generation, experimental design, execution, and peer review
- Full research loop automation: agents propose experiments, validate hypotheses against literature, pivot when evidence contradicts their assumptions, and document failed approaches alongside successful ones
- Cockpit interface for defining research problems, monitoring live experiment fleets, and reviewing leaderboard progress with per-claim evidence tracing
- Built-in experiment management: parallel research arms, automated GPU dispatch, experiment time capping, and monitoring agents that terminate poor-performing runs early
- Proven capability to reproduce and improve published baselines, including fixing bugs in existing codebases and transferring techniques across domains (e.g., Muon optimizer from Sudoku to ARC-AGI)
- Open-source components (configgle, sagent, Trackinizer) and Apache 2.0 licensing for discoveries, with published case studies showing 984 experiments and 882,534 lines of scientific discourse generated in one month