What Fulcrum does
Fulcrum is building an agentic debugger for AI systems. The Y Combinator-backed company makes software that diagnoses why AI agents fail and uncovers bugs in the environments they run in, with a focus on agents and reinforcement learning environments.
Key capabilities
Fulcrum's monitoring identifies why agents do not succeed, surfaces bugs in environments, and detects fake solutions, reward hacking, and catastrophic failures. Its red-teaming agents run experiments that plug into environments, agent traces, and agent source code to find the root cause of issues. Investigations are turned into explorable reports that users can chat with to understand what went wrong.
Who it's for
Fulcrum is aimed at teams building RL environments or deploying AI agents who need to debug complex, hard-to-reproduce failures. The founding team met at MIT doing research on large language models, has published at conferences including NeurIPS and CoLM, and has built widely used open-source software. By making agent failures observable and explainable, Fulcrum helps developers trust and improve the AI systems they ship.