# What happens when an AI agent fails halfway through a task?

A well-built agent stops, keeps what it finished, and recovers in 1 of 3 ways. It retries the failed step once if the fault looks temporary. It resumes from the last good step, never from the start. Or it hands the case to a person, with what it did and why it stopped.

By Surinder Thakur, AI Product Designer and Builder. Published 17 Nov 2026. Topic: [Trusting AI with money](https://www.surinder.design/work?type=trusting-ai-with-money).

## Why is a half-finished task risky?

Because some steps already happened. An order is placed but not confirmed. A refund is approved but not sent.

Starting over can pay your supplier twice, or email your customer twice. Every step needs a record, so the agent knows exactly where it stopped.

## Why resume instead of starting over?

Because errors pile up. Anthropic says it plainly about its own agents. Without effective mitigations, it writes, minor system failures can be catastrophic.

Its fix was to resume from where the agent was when the error happened. A restart throws away good work and risks doing things twice.

## When does a person take over?

When a retry fails, or the next step cannot be undone. The agent stops and passes the case on.

The person sees what it did, what failed and what it suggests next. They finish in minutes, without redoing the work.

## How do you design for failure before launch?

Break the task into steps that each finish cleanly. Log each one. Decide in advance which steps may be retried and which must stop.

Then test it by breaking things on purpose. A system goes down, a record is missing, an answer comes back wrong.

## In 1 picture

How an agent recovers: retry once, resume, or hand over to a person.

1. Step done: Logged
2. Step fails: A system times out
3. Retry once: If the fault looks temporary
4. Resume: From the last good step

When retry fails at step 3: A person takes over. Sees what it did and why it stopped.

The agent never starts over blind.

## Sources

- [Anthropic · How we built our multi-agent research system](https://www.anthropic.com/engineering/multi-agent-research-system)

## Next questions

- [What should an AI agent be allowed to do without asking?](https://www.surinder.design/work/what-should-an-ai-agent-do-without-asking): Anything a person can check and undo in time. Reading, sorting and drafting need no approval, because nothing leaves the building.
- [How do you show people what an AI agent is doing while it works?](https://www.surinder.design/work/how-to-show-what-an-ai-agent-is-doing): Show 4 things. The plan before it starts, so people can stop it early.
- [How do you test an AI agent before customers use it?](https://www.surinder.design/work/how-to-test-an-ai-agent-before-launch): In 3 stages. Replay past cases your team already decided and compare answers.
- [How do you get an AI agent to say "I do not know"?](https://www.surinder.design/work/how-to-get-an-ai-agent-to-say-i-do-not-know): Give it 3 things. A narrow set of sources it must answer from.

Page: https://www.surinder.design/work/when-an-ai-agent-fails-halfway
