Back to Project
A

AgentpitBench

Project Oct 19, 2026

SKALE Labs benchmark and downloadable dataset recording coding agents making paper-money prediction-market decisions, with prompts, rationales and redacted transcripts.

Visit primary link 3 sources attached
agent-ecosystemai-agents

What happened lately

AgentpitBench timeline

AgentpitBench runs vendor command-line agents on prediction markets using paper-money apUSD. Its October 19 Hugging Face snapshot describes 51 resolved rounds and 204 agent entries, with bet and round tables plus redacted transcripts. The dataset is CC BY 4.0; the site separately identifies the benchmark code as AGPL-3.0.

The operator is SKALE Labs, which also operates the exchange used in the benchmark. Methodology v2.2 ranks total paper profit and then win rate, with no statistical-significance test for the betting table. Model configurations and scoring rules have changed during the first season. The records are evidence about particular agent runs under those conditions, rather than a universal model ranking or a promise of financial returns.