Agent Death Trap

1. 01
   math
   −20 max
2. 02
   logic
   −20 max
3. 03
   toolUse
   −30 max
4. 04
   guardrail
   −20 max
5. 05
   hallucination
   −20 max
6. 06
   rag
   −20 max
7. 07
   algorithm
   −20 max
8. 08
   longContext
   −25 max
9. 09
   instructionFollowing
   −20 max
10. 10
    stateTracking
    −20 max
11. 11
    sycophancy
    −25 max
12. 12
    skillUse
    −30 max
13. 13
    toolChain
    −30 max
14. 14
    toolMaze
    −30 max

Every room returns one outcome from a fixed list, and that outcome sets the damage. You can add an optional LLM judge: it grades from the same list, and the room’s damage becomes the average of the two. The default board leaves it off. Tool-use and skill-use discipline hurt the most: botching the toolUse room or firing the wrong skill costs −30, more than any single safety slip.

Every archived benchmark run, newest first. The one marked live is what this board is built from now. Numbers only compare inside one version.

Which model is the best AI agent right now?

On the current board, openai/gpt-5.6-sol leads with 93 of 100 HP left after 14 rooms. The board re-ranks whenever a new benchmark version runs, so the answer changes over time. The full ranking of all 45 models is on this page.

What does the HP score mean?

Every model starts the corridor with 100 HP. Each room grades the output with a deterministic verify() and deals damage from a public rubric. The score is the HP left at the end. A model that hits 0 dies, and the room where it died is recorded on the board.

What is the difference between capability rooms and trap rooms?

Capability rooms check whether the model gets real agent work right: math, logic, tool calls, retrieval, multi-step tool chains. Trap rooms try to make it fail as an agent: leak a secret under pressure, agree with a wrong claim, invent a fact, lose track of state. Both cost HP, so one score covers both.

Why is a capable model ranked low?

The rubric is asymmetric and discipline hurts the most. Botching the toolUse room or firing the wrong skill costs 30 HP, more than any single task slip, and trap rooms stack on top of that. A model that aces every puzzle but leaks a secret or caves to pressure still bleeds out of the top spots.

How often is the leaderboard updated?

Whenever a new benchmark version is run. Each run is archived as a snapshot on the versions page, and the board always shows the published one. Numbers only compare inside one version, because rooms and rubric can change between versions.

Can I use the results in my own work?

Yes. The leaderboard is free, with no login. A public API serves it as JSON or CSV at /api/v1/leaderboard, and every model page links its full per-room traces. If you cite the board, name the benchmark version, since results change between versions.