
























What if we could observe how AI agents follow agentic skills? Wait, we can - automatically, by spawning another agents processes in background and reading its output continuously.
What if we could observe how AI agents follow agentic skills? We can. Spawn another agent process in background, read its output, score against ground truth. TDD or User Testing for skill files.
I was updating our serpapi-web-search skill and wanted to know: does this specific documentation help LLM agents correctly solve the task?
This approach helped self-fixing several documentation issues:
SKILL.mdExample prompt: "How many Google Maps reviews does Tartine Bakery have?"
Ground truth: 5,915.
With the skill: agent calls SerpApi, gets 5,915.
Without: Opus refused, Fable launched Playwright (works 5/12 times), GPT guessed confidently wrong.
For the different task, GPT gets exact flight prices WITHOUT the skill ($299, matching API answer, 5/6 correct). Fable and Opus: 0/12.
We ran 432 of these.
| Task | With | Without | Notes |
|---|---|---|---|
| Scholar citations | 100% | 28% | Without gives stale 265K vs exact 257K |
| Flight prices | 100% | 14% | GPT gets it via web search; Fable/Opus can't |
| Maps reviews | 100% | 19% | Fable uses Playwright; Opus refuses; GPT guesses |
| Hotel prices | 100% | 0% | Refuses: "I can't look up live hotel prices" |
| Image source | 100% | 0% | Hedges or guesses wrong domain |
| Product price | 100% | 33% | Gives $449-499 estimates; actual $649 |
| Shopping price | 67% | 0% | 1 auth failure; gives ~$198 vs $248 actual |
| YouTube views | 100% | 92% | Zero lift. Web search returns exact counts |
| App Store rating | 100% | 100% | Zero lift. Public API endpoint |
| JWST mass | 97% | 97% | Control. Training data |
| News headlines | 100% | 100% | Control. Any web search works |
Scholar + Flights + Maps: WITH 100%, WITHOUT 20%. Haiku provided correct answers only when skill was provided.
The AUT (Agent Usability Testing) skill runs on two conditions:
WITH scores higher: the skill was useful. WITHOUT scores same: it wasn't useful.
It guides an agent through the steps:
I did several mistakes during evaluation:
$HOME (all configs and extensions) vs stripped HOME. Not a fair comparison.serpapi binary in ~/.cargo/bin survived PATH shadowing.
I ran a third condition: tool on PATH and authenticated, but no skill file.
0 out of 24 runs. Never typed serpapi. Never explored PATH. Never discovered it.
We tested this "TDD for skills" approach with Copilot CLI and SerpApi and fixed several documentation issues, that the /review on clean context didn't notice. Please give the it try and share your results.
P.S. For transparency, below are all tasks during the tests:
此内容由惯性聚合(RSS阅读器)自动聚合整理,仅供阅读参考。 原文来自 — 版权归原作者所有。