Fulcrum: The agentic debugger for AI systems

The agentic debugger for AI systems

We build red-teaming agents that help developers fix their RL environments and improve their agents. Our customers are tired of being bullshitted by their agents.

Active Founders

Kaivalya Hariharan

Founder

MIT CS & Math ’24. MEng AI candidate under Jacob Andreas (LINGO) group. Best Paper—NeurIPS ML Safety ’22; follow-up at NeurIPS ’23 and benchmark paper accepted to NeurIPS ’24. Emergent Ventures Fellow ’25. Founder of MAIA (MIT AI Alignment). Research stints at Redwood Research and Truthful AI. USABO Nationals finalist (camper), two-time AIME qualifier.

Kaivalya Hariharan

Founder

MIT CS & Math ’24. MEng AI candidate under Jacob Andreas (LINGO) group. Best Paper—NeurIPS ML Safety ’22; follow-up at NeurIPS ’23 and benchmark paper accepted to NeurIPS ’24. Emergent Ventures Fellow ’25. Founder of MAIA (MIT AI Alignment). Research stints at Redwood Research and Truthful AI. USABO Nationals finalist (camper), two-time AIME qualifier.

Uzay Girit

Founder

working on better human AI collaboration - building oversight interfaces for AI agents. I did ML research on in context learning and scaling laws, studied math and CS at MIT, into writing and climbing. From Paris. Website: https://uzpg.me

Uzay Girit

Founder

working on better human AI collaboration - building oversight interfaces for AI agents. I did ML research on in context learning and scaling laws, studied math and CS at MIT, into writing and climbing. From Paris. Website: https://uzpg.me

Company Launches

Fulcrum: The Agentic Debugger for AI Systems

See original launch post

TL;DR:

We’re building the best software for debugging AI agents and their environments. If you’re building RL environments or deploying agents, get in touch with us and try out our product.

https://www.youtube.com/watch?v=WXa-cGlg6E4

The problem
Agents are incredibly complicated. Evals are a mess. We need to understand what our agents are capable of and if our RL environments even work.

The solution

Our monitoring system identifies why your agents don't succeed, uncovers bugs in your environment, and detects fake solutions, reward hacking, or catastrophic agent failures. Our redteaming agents run experiments to diagnose your agent issues. They plug into your environments, agent traces, and agent source code to find the root cause of your issues. Finally we take all of these investigations, and turn them into intuitive, explorable reports you can chat with.

Who we are
We met at MIT where we were doing research on LLMs. In our latest paper we built a method for scalably generating software environments of arbitrary difficulty. Doing it ourselves gave us the know-how to build a great environment debugger. We have published at top ML conferences (NeurIPS, CoLM) and built widely-used open source software (5k stars, >100k downloads). Our team has won a total of four medals at international olympiads in physics and economics.

Call to action:

Email us at team@fulcrumresearch.ai if you’re building environments or deploying agents, and share this around with people who are. Or chat with us!