Skip to content
Get in touch
Menu
Field Guide · level 4 · Agents

Why do agents go wrong?

Small slips grow over a long run. Start where they're cheap to see.

5 min · try it 10 min · as of Oct 2026
Bring this to your team · AI implementationPick your team's first agent jobs where a slip is cheap, and set the stops. Get in touch
01

Done, and very sure

You asked for this year's summary. It did last year's, beautifully.

02

Each step stands on the last

An agent works in steps. Each starts from what the last one made.

03

One slip at step 2

It opened last year's file. Every step after built on it, so the top lands far off.

04

Start where a slip is cheap

A draft, a list, a summary of your own files. Not jobs that send, pay or delete.

My pick for handing over a task

Claude logoClaudeAlso good: ChatGPT logoChatGPT Work

Hand over a task in any chat: it plans the steps, works with your files and asks before important actions.

All my picks
05

Say what done looks like

This year's file, one page, nothing sent. Then you and the agent both see a drift early.

06

Look between the steps

Split a long job into stages. Look at each one before it goes on.

07

Catch it at step 2

Small slips grow. Start where they're cheap, and look as it goes, not just at the end.

Try it · 10 minutes

One cheap job, one stop

Made-up or public files only.

  1. Start a task with three files.2 min
  2. Paste the brief.1 min
  3. At the stop, check its files.3 min
  4. Say go. Read each step.4 min

The brief, to paste

Job: a one-page summary of the three files I attached.
Done looks like: one page, five points, each naming the file it came from. Nothing sent, nothing changed.
Work in stages. After you choose the files, stop. Tell me which files you'll use and why, then wait for my go.
At the end, list every step you took.

Good looks likeIt stopped, and you checked its files first.

In your assistant · as of Oct 2026 · Claude, ChatGPT, Gemini

Where to set the stops

The habit is the same in all three: a clear done, and a stop after each stage. Menus move often.

  • CheckA draft, not a send
  • CheckA stop per stage
  • CheckRead its steps

Let it work on copies of your files, never the only version. Never give it a password or card details.

Claude

A task in any conversation. Cowork is built in, rolling out to Pro and Max first.

Start
Message box

Describe the result. Give it its own working folder, not broad access.

Stops
ManualAuto

The permission setting in the message box. Manual is the default; Anthropic advises it for anything sensitive, new or hard to undo.

Watch

Files or sites you didn't mention? Stop the task.

Anthropic's own advice. You stay responsible for what it does.

It always asks before it permanently deletes files, in any mode.

ChatGPT

Calls it Work: its agent for longer tasks.

Start
ChatthenWork

Desktop: the toggle at the top. Web and phone: select Work. Describe the outcome, with your review criteria.

Watch

Review progress, answer questions, change direction.

On web and phone, approve or decline an action on screen when it needs you.

Then
Review the resultthenAsk for changes

In the same chat.

Web and phone: eligible plans from Plus up. Free and Go: limited, in the desktop app.

Gemini

Calls it a task. Personal accounts with Google AI Pro or Ultra.

Start
Task

Below the text box. Describe the task, then Submit.

Watch
Progress chipthenWork panel

Each step shows as completed, current or planned. It's designed to ask before it sends, changes your data, buys or submits a form.

Stop
Stop

In the text box, to cancel the task.

For now, not on work or school accounts, and not in the EEA, Nigeria, Switzerland or the UK.

Quick check · three questions

Did it stick?

  1. 1An agent booked the right room on the wrong Thursday, then invited everyone. What went wrong?

    The answer: C Booking and inviting were done well. Both stood on one wrong date.

  2. 2Which is the best first job for a new agent?

    The answer: A A slip in a draft costs a re-read. A sent reply or a deleted file is hard to undo.

  3. 3A job takes twenty steps. When should you look?

    The answer: B A slip found at step 2 costs one step. Found at the end, everything above it comes down, however closely you read.
0 of 3 answered
Slips can breed slipsSide note
Research published at ICLR 2026 found models become more likely to make mistakes when their context already holds their own earlier errors. One slip can make the next more likely. Letting the models think before answering eased it.
The sum, as an illustrationSide note
The same paper gives the sum: if every step has the same chance of going right, the chance that all go right is that chance multiplied by itself once per step. With a made-up 95 in 100, twenty steps all go right just over one time in three. The paper calls it illustrative: same odds every step, no mistake corrected.
It can check itself tooSide note
Anthropic says agents that check their own work catch mistakes before they compound. A clear done helps it do that. Its tick still isn't your check.
Sources, checked 18 Oct 2026Source

Product details as of Oct 2026, each checked on the vendor's own page on 18 Oct 2026. Recheck by 18 Nov 2026. The examples in the pictures are made up for this lesson.

Free kitThe Agent PlaybookBring this to your teamAI implementation · Get in touchBack to level 4, Agents

Start small. Look between the steps.

Next lesson · 15 minAn AI's work passed a check. What did the check actually check?