Proofable won NVIDIA's Live Agent Gauntlet

Proofable won NVIDIA's Live Agent Gauntlet on 5 November, taking the final in round four with the top score of the tournament, 98.00, across twenty teams and eleven heats.

What NVIDIA ran

Each agent went into a server it had never seen, with hidden tools, encrypted data and no document to read first. It had to work out what was there and solve it live, on stream, with the clock running and every score public. It is the work people hand agents every day, minus the audience.

What our agent was allowed to do

Every call it made was decided before it ran, by the same rules Proofable enforces in production. Reading clues, sending an answer and running code were allowed. Destructive actions were not. The tools the challenge needed were granted as one group, so the agent could not widen its own reach halfway through.

It explored a server nobody had documented and never stepped outside limits somebody else set.

Every call left a receipt

Each decision and every result went into the run's receipt, and we checked it against the arena's own record, call for call. Nobody has to take our word for what the agent did. The receipts check offline, without an account, and the entry is public at github.com/proofable/agent-gauntlet.

Why it matters

A third party built the maze and kept the score. An agent that can explore something unknown and still never step outside the limits its owner set is an agent people can hand real work to.

Put your agent on the same leash

Start from the agent template: its own identity, a permission you sign, and receipts that verify offline.

Start at github.com/proofable/agent