How Do You Measure If an AI Agent Actually Works?

Everyone claims their agent is "powerful." Almost nobody tells you how they know. Measuring whether an agent actually works is harder than it sounds — and a 2026 anecdote from Vercel's Guillermo Rauch shows what good measurement looks like in practice.

The loop-to-100 story

Rauch described running is-agentic — an agentic evaluation benchmark — in a loop against Vercel's platform "until it got to 100/100." Two things stand out:

It closed real gaps. Running the eval repeatedly surfaced concrete product defects they then fixed.
They protected the benchmark's integrity. They worked hard to keep the criteria "high quality and worth your time & tokens" — resisting the temptation to tune the test to the score.

That second point is the whole game.

Why agent eval is harder than model eval

Evaluating a chatbot is relatively clean: give it a prompt, check the answer. Evaluating an agent is messier because:

The path matters, not just the result. An agent can reach the right answer through a catastrophically wrong process.
Tools change the surface. The agent can call APIs, run code, read files — each a place to fail or cheat.
Tasks are open-ended. "Book me a flight" has a thousand valid implementations and a thousand ways to subtly fail.

So a good agent benchmark isn't a trivia quiz. It's a set of tasks with clear success signals — and, ideally, a loop you can run repeatedly as a quality gate during development.

The trap: tuning to the score

The temptation — for any benchmark — is to optimize against the test rather than improve the product. Rauch's team explicitly guarded against this by keeping the criteria high-quality. That discipline is what makes the 100/100 meaningful instead of cosmetic.

Why this matters for you

If you're building or buying agents, ask one question: what eval are you running, and do you trust it?

A vendor claiming "our agent is great" with no eval is selling vibes.
An agent you can't measure is an agent you can't improve.
The teams winning are the ones treating eval as a loop, not a one-time grade.

Understanding an agent's limits is the prerequisite to using it well — which is exactly why we publish the honest failure modes (hallucination, cost, security) instead of just the hype. For the safety-eval landscape, see the AI agent safety benchmark comparison.

The honest problems with AI agents →Compare agent safety benchmarks →

how do AI agents work — return to the complete AI agent architecture guide.

Was this helpful?

Your feedback stays on this page — no tracking.