Armory
Source
Browse
Evals

agentbenchOurs

Use to put a number on harness quality: run an agent harness against a task set and get a score, so harness changes are validated by evidence. It is the eval backbone of a self-improving loop.

Score
Unranked
Evidence
No signals yet
Last commit
as last read from GitHub; most reads are from 2 Sep 2026 or later
Listed

Install

No one-command install. Set it up from its source.

What it is

A benchmarking harness that scores an agent harness against a task set, producing a single comparable quality number. It is the measurement instrument for the "same model, different scaffold" thesis: the way you prove a scaffold change moved the needle (e.g. a jump from a poor score to a strong one) rather than just claiming it did.

When to use it

Whenever you change a harness and need to know if it improved or regressed, and as the scoring step in any automated self-improvement loop. The trigger is "score this harness."

How to install / invoke

Run the benchmark CLI against a harness with a defined task set and record the score. Track scores across versions to watch the harness climb (or catch a regression early).

Notes

A score is only as good as the rubric and tasks behind it, so pair the benchmark with a clear rubric and a stable set of golden tasks. This is the same tool listed under CLIs; here it is framed as the eval primitive.

Related

Not Indexed

  • harness-rubric
  • golden-tasks
  • autolab-nightly
  • harness-review