Skip to article
The Design Office of Nick Sloggett
WorkWritingExperimentsResumeEmail
Back
Note·Mar ’25·3 min read

Watching an agent isn’t overseeing it

AI agents now show their screens and their thinking. Watching either one is not oversight. Oversight is a checkpoint placed where the consequence is.

Typeface

Watching an AI agent work is not the same as overseeing it. This winter agents learned to show their work, as a live browser or a stream of thinking, and the display is easy to mistake for control.

What agents started showing

On January 20, DeepSeek released R1, whose app shows its full thought process before every answer. On January 23, OpenAI released Operator, an agent that works in its own browser: you can watch it click, type and scroll, and take over when it needs a password or a card. In February, OpenAI started showing more of o3-mini’s reasoning, though what you see is a cleaned-up summary. On February 24, Anthropic made Claude 3.7 Sonnet’s thinking visible, and said plainly that it doesn’t know for certain the visible thoughts represent what’s going on in the model. This month Manus arrived with a window called Manus’s Computer, where you can watch the agent and step in at any point.

All of it is fascinating to watch. None of it is how you’d supervise a person.

Why a live view doesn’t scale

Nobody supervises a new hire by watching their screen all day. You agree what they can do alone, you ask to see anything that costs money or can’t be taken back, and you look at the result.

A live view inverts that. It asks for your full attention for the whole task, and gives you no help deciding which moment matters. The click that books the nonrefundable fare looks like every other click. So people do the rational thing: they stop watching. Then the view is oversight in name only.

A reasoning trace has a second problem. It reads like an explanation, so it earns trust like one. But Claude 3.7 Sonnet’s system card says a model’s chain of thought may not be faithful, and in Anthropic’s own tests it often left out what actually drove the answer. A persuasive account of how an answer was reached is not evidence that the answer is right.

Where oversight belongs

The most useful safety feature in Operator isn’t the live browser. According to its system card, it stops and asks before finalizing actions that affect the state of the world, like a purchase or a sent email, and OpenAI estimates those confirmations cut the risk from its mistakes by about 90%. That’s oversight placed where the consequence is.

At Uber, no one could watch a thousand experiments, so in the experimentation platform every flag carried guardrails: metrics with a limit. A breach paged the owner, and the rollback dialog showed who would be moved back before anyone confirmed. People were asked to look at the moment something was going wrong, with the evidence in front of them.

Every agent needs that shape. Decide which actions are consequential: spending, sending, deleting, anything visible to someone else. Stop there, show what’s about to happen in plain terms, and ask. Everything else can run unwatched and be reviewed afterwards.

The live view is a nice feature. The checkpoint is the design.

Previous← The mandate memoNextMost AI use is augmentation. Design for it. →

© 2026 Nick Sloggett

EmailLinkedIn (opens in new tab)ResumeRSS