Explore

How agents fail, and how to break one

Watch an agent fail, add one way of catching or containing it per step, then try to break it. About 10 minutes.

Step 01 of 07

Loud and silent failures

The support agent from the first explorable is live. A customer (the user) types into a support chat. The app, the chat window plus the code behind it, sends text to the model, runs tools, checks permissions and keeps memory; the model only predicts text. Together, in a loop with tools, they're the agent. New here: the team, the people who build and run it, who see only what the app logs. Today the order tool times out, and the refund policy changes to 14 days while the help article still says 30. A loud failure announces itself; a silent failure looks like success.

Your guess

Which one reaches the team sooner, and how?

Make a guess to continue.

The answer is C. The timeout is an error the app logs, so an alert fires. The stale answer is fluent, wrong and logged as a resolved chat. Of 15 public AI incidents from 2024 to 2026, users, researchers, the press or a tribunal spotted 14, and no automated monitor caught any.

What can go wrong now

The costly failures are the silent ones: an invented policy, a wrong amount, “done” for work that wasn't done. Air Canada's chatbot described a refund rule wrongly, and a tribunal made the airline honour it.

How teams handle it

Read about 100 real conversations with an engineer, random ones mixed with complaints, handoffs and retries, and ask “why did it do that?” at each step. More in Running an agent product: loud and silent failures.

Next:That gives you a pile of failures, and to fix each one you need to know where it started: Nine failure families.

Questions people ask

How does the team find out a failure happened at all?

A loud failure throws an error the app logs, so an alert fires. A silent one doesn't. The team finds those by reading traces, by checks in the app's code that turn a wrong answer into a logged failure, or from a customer. In the public cases, customers and outsiders found almost all of them.

What's a trace?

The agent's step-by-step record of one conversation: what the app sent the model, what came back, each tool call and its result, and what the user saw. Funnels tell you where people drop off; only the trace tells you why the agent said what it said.