Skip to main content
Evals answer a question you otherwise answer by guessing: is this version actually better than that one? Pick a prompt, skill or agent, pin up to three variants of it, run them all on the same input, and read the results next to each other.

Opening evals

The top bar carries an Editor / Evaluation toggle. Switching to Evaluation opens a drawer beneath the editor area. It follows whatever you’re looking at, so switching tabs re-targets the eval.

Setting one up

1

Pick a target type

Prompt, Skill or Agent — one type per eval. You compare a thing against itself, not a prompt against an agent.
2

Add variants

Each variant is a version and a model. Up to three. Any model active in the Space is available.
3

Fill in the input

Prompt variables come from the prompt’s own values. A required variable left blank is an error rather than an empty run.
4

Run

All variants run in parallel against the same input.

Reading the results

Outputs sit side by side, with tokens and cost per variant. They’re shown as outputs, not as a diff. Diffing three model outputs is noise — nearly every token differs even when the substance is identical. Read them; the interesting differences are rarely textual. A variant that fails doesn’t take the others down. It reports its own error and the rest keep going.

Cost

Before running, you see an estimate. Three variants means three real runs charged the normal way, so a three-variant eval costs roughly three times a single run.
Skill versions can’t be pinned. Skills are filesystem-canonical — the file on disk is the skill — so there’s no version history to select from. Compare skills across models instead.

History

Every eval is kept and can be reopened. Results are written as files in the Space, so they travel with it and can be committed like anything else. Delete one when you don’t need it.

Run an agent locally

Single runs, for when you’re not comparing.