Agnt Bench

by AGNTOfficialPlugin

AGNTBench P0 — benchmarks AI models on the things they actually do inside AGNT. Tracks: T1 schema fidelity (typed tool args), T3 grounded editing (read-before-edit), T9 shell dialect (host-shell correctness). Every task is derived from measured production failure signatures. Scoring is fully deterministic — no LLM judge.