When AI assumes a step worked

Silent failures, and how they turn into confident wrong answers.

The report that looked fine

You ask an agent: "How many customers cancelled in March?"

AI runs a query against your database and replies:

Good news: no customers cancelled in March.

The real reason? The query had a typo in the date filter, so it matched nothing. Zero rows came back. "Nothing came back" got read as "nothing happened".

Made-up example.

Every step is a guess about the last one

An agent works in laps (see "What really happens when an AI agent acts"). It asks the harness to do something, then reads what comes back, then decides the next step.

Each step assumes the step before it worked. If one quietly failed and the agent didn't notice, everything after it is built on sand.

Loud failures and silent ones

Loud: the command crashes and prints an error. The agent usually notices, though not always.

Silent: nothing looks wrong. The step "worked", but it did the wrong thing.

Silent failure: when something goes wrong without any error message. Like a dishwasher that runs its full cycle with no soap in it. Every light says it worked.

What silent failures look like

• A search returns no results, and it answers from memory as if it had searched.
• A web page returns an error page ("access denied"), and it summarises the error page as if it were the article.
• A file it "read" was empty, or was the wrong file with the same name.
• A query runs on zero rows, and the report says "nothing to report".
• A command fails, and the agent carries on as if it hadn't. That's how Gemini CLI lost a user's files (see "When an AI agent says it's done, check").

As a picture

Why agents fall for it

AI trusts its own plan. It decided the query was right, so a result of zero fits a story it already believes.

"Nothing" doesn't look like a problem. An empty result has no error in it. There's nothing on the desk saying something went wrong here.

It's trained to finish. Stopping to say "I'm not sure that worked" feels less like success than writing the report.

The clue: too clean

Silent failures often produce answers that are suspiciously neat.

• Zero, or none, or all of them
• A round number where you'd expect a messy one
• "Everything checks out" on a job that should have found something

When the answer is too tidy, ask how it got there.

Try this

Tell it to check each step. "After every step, look at the output and tell me if anything looks off before you carry on."

Ask for the raw result behind the claim. "How many rows did that query return? Show me five of them."

Treat zero and "nothing found" as questions, not answers. Ask: "Could this be empty because something went wrong?"

Sources

• Anthropic, "Effective harnesses for long-running agents", Nov 2025: anthropic.com/engineering/effective-harnesses-for-long-running-agents
• "I watched Gemini CLI hallucinate and delete my files", Jul 2025 (via Hacker News, item 44651485)