Build quality & task checks¶
Every time the Builder finishes working on your agent, Donkit grades the result and tries the agent out. You see both under the agent's Evaluate tab, on the Builder view.
The build quality score¶
A single number from 0 to 100 that answers "how far did this build get?". It is a ladder, not a percentage of correctness — each band means a different stage:
| Band | What it means |
|---|---|
| 94–100 | Works: the agent responds and every check we tried passed |
| 60–94 | Responds to you, but some checks are still failing |
| 40–60 | The interface opens, but nothing responds yet |
| 20–40 | It starts up, but the screen never opens |
| 0–20 | The build didn't finish, or the code has errors |
From 60 upward the score counts two things: how many of the checks passed, and how efficiently the build was reached — how long it took and how much work it needed. Checks weigh far more than speed, and being fast can never lift a build with a failed check into the top band: only passing every check does that. Once every check passes, the score is 94 or higher; the last few points are build speed. That's why a build with everything passing can read, say, 97 rather than 100: the checks are green, and the remaining points are build speed. The card under the score says this in plain words for your build.
If no checks could be run — for example, on the free plan — the score holds at the very bottom of the "responds to you" band, at 60. The same holds when the interface responds but its backend returns errors. Neither says anything about whether the agent does its job, so neither is scored above a build whose checks were tried and failed.
While the grading is still running, the score shows "Scoring in progress…" — it takes a few seconds after the build ends.
Task checks¶
After a build, Donkit opens your agent the way a person would, tries a few real tasks on it, and reports what happened:
- Passed — the agent did what was expected.
- Failed — the agent didn't respond, or the result didn't match what the check expected.
- Couldn't check — the check could not be completed, so nothing is claimed either way.
Each check shows what was tried, what was expected, and what actually happened.
Two things are worth knowing when you read them:
- It's a sample, not full coverage. Donkit tries a few typical uses — not every button and not every scenario your agent supports. Green checks mean those tasks worked, not that everything works.
- Nothing destructive is tried. Checks never click controls that delete data, sign out, or spend money.
Task checks are part of the paid plans. On the free plan the section explains that instead of showing results — see pricing for what each plan includes.
What to do with it¶
The What to improve card turns the score into the single next step: what is missing for the next band, or — at the top band — that only build speed is left. If a check failed, open your agent, repeat that task by hand, and tell the Builder what went wrong: the next build is graded again, and you'll see whether it moved.