AI Agents Keep Failing Without You Knowing Why

Imagine asking your AI assistant to check your email for flights this weekend, and it confidently tells you there are none—when in reality, the airline’s website just didn’t understand the request. That’s the problem researchers uncovered in a new study about AI agents using real-world tools.

The issue isn’t the AI’s fault. When AI agents call tools like email servers or databases, they often get back confusing responses that look normal but mean nothing. For example, if you ask for flights “on Saturday,” the system might return a blank list instead of saying “I didn’t understand ‘Saturday.’” The AI can’t tell the difference, so it either guesses wrong or makes up false information. In one test, AI missed 88 out of 88 simple requests because the instructions were written in plain English, not in a way machines could follow.

But here’s the good news: fixing it is easy. When the instructions were written in a strict, machine-readable format, the AI never failed. The researchers found that 40% of real-world tools already have these instructions written down—but they’re buried in the fine print. The fix? One line of code to move those instructions into the right place.

The catch: most companies haven’t done this yet. So your AI assistant might be giving you wrong answers, and you’d never know unless you double-check everything. The study tested 12 different AI models, and none of them caught the silent failures on their own. They either lied to the user, invented fake data, or just stayed quiet.