2026-10-31 · America/Los_Angeles · 社区动态 · #15

I built an agent benchmark tool to read how agents actually solve problems, not just the score

I'm curious how agents get better over updates, not just the score but how they really solve problems, so I built hbench. I want to read agent trajectories closely, and a final score doesn't tell me much. Two agents can get the same pass rate in different ways. What it tried, where it went in circles, and which tool call changed the outcome. Here is an example: DeepSeek's harness and its model tried to pass by reading the solution from benchmark's public sources. It got reward 1.0, so in the results table it's just a normal pass. https://github.com/node-q/hbench

热度 45.0 / 100；排名与评分保留该期记录。