Open-source benchmark targets the tool layer: auth, retries, memory failures that agent evals miss

Kind-Atmosphere9655 · reddit · 2026-09-04

A Reddit developer argues most agent evals only test whether the model picks the right answer, missing a whole class of failures: the model chose correctly, but the tool layer used the wrong account, over-requested permissions, duplicated side effects on retry, claimed success without verifying final state, or lost memory.

He's building an open-source benchmark for that layer, holding agent, model, prompt, and task fixed while swapping only the capability provider, with a separate verifier checking external state. The repo is pre-alpha: 10 public task contracts plus a working verifier and harness, but no production backends or official provider scores yet.

He's asking the community for the nastiest real-world failures in auth, tool calls, memory, approvals, retries, or sandboxing—and what evidence would convince them a task actually succeeded.

Original post →

More from coding & agent

coding & agent channel →