Agent Death Trap

Founder’s note

Why I Built Agent Death Trap I built Agent Death Trap because choosing an LLM for an AI agent is surprisingly difficult. Most benchmarks tell us which model is better at coding, math, reasoning, or knowledge. Those numbers are useful, but when I started building real agents, I realized they don’t answer the question I actually care about: Which model can I trust to run an agent? An agent has to do much more than generate a good answer. It needs to understand tool contracts, call the right tools, follow multi-step workflows, keep track of state, retrieve the right information, resist bad instructions, avoid hallucinating, and recover when things don’t go as planned. A model can be excellent on traditional benchmarks and still fail badly somewhere in that chain. So I built Agent Death Trap as a different kind of benchmark. Instead of giving models isolated tests and producing another abstract score, every model goes through the same corridor of challenges. Each room represents either something an agent needs to do well or something that commonly causes agents to fail in real systems. Every model starts with 100 HP. Mistakes cause damage. Serious agent failures cause more dama