I built a small tool to test my own AI agent in 10 minutes, instead of trusting that it "looked right."
One test from it: "Test a clamp function." Write tests for clamp(value, min, max) — a value inside the range, below it, above it, and both exact boundaries. Five cases, with expected results.
I ran two answers through it:
- A full answer — clamp(5,0,10)=5 inside; clamp(-1,0,10)=0 below; clamp(11,0,10)=10 above; clamp(0,0,10)=0 min boundary; clamp(10,0,10)=10 max boundary — scored 4/4. What-to-fix: "No fix needed for this task."
- A short answer that only tested inside, below, and above — no boundaries — also got 4/4 from the auto-check, because the only keyword rule is "clamp." That's the honest catch: the auto-score suggests, you confirm. The task check says "covers all five cases," so a human reviewer unticks "Is correct" for the answer that skipped the edges — and that's exactly where clamp bugs live.
That is the whole point of the tool. It scores 4 things per task: finished, correct, did not make things up, and took no wrong action. Auto-score suggests, you confirm — so you get speed without handing the judgment to a black box.
The tool has 20 tasks: 10 developer, 5 personal-life, and 5 creator. It is free, with no API key, no login, and nothing leaves the page. It is a quick first check, not a CI or production replacement — use it first, then graduate to heavier tools when you need them.
Try this one task yourself: https://jefferson999.github.io/agent-reliability--score/
Paste your agent's answer, see the score, and read what to fix. If your agent passes the boundaries, try the other 19.
Top comments (0)