Agent Mindset
← back to notebook

Basics

Delegate, Then Verify

The core habit of working with an AI agent: hand off the work, but own the outcome.

7 min read

An agent is a collaborator, not a vending machine

When you use a search engine, you operate a tool: you type, it returns links, and you do the rest. An AI agent is different. You delegate a task to it — "add a dark-mode toggle to this page" — and it plans, takes steps, and comes back with a result.

The useful mental model is a fast, knowledgeable, eager junior collaborator. It can do a surprising amount on its own, but it sometimes misunderstands, makes confident mistakes, or does more than you asked. So you direct it and check its work — just like you would with a new teammate.

You delegate the work, but you own the outcome

Delegating does not mean walking away. The agent can produce something that looks finished and still be wrong — a feature that doesn't handle the real case, or a summary of a document that invented a detail.

The habit that separates effective agent users from frustrated ones is simple: specify clearly, then verify the result yourself. Don't accept "Done!" as proof. The agent did the work; you are accountable for whether it's actually right.

From the field

SW

Simon Willison

Creator of Datasette; long-time open-source developer who writes extensively about coding with LLMs

software
A computer can never be held accountable. That's your job as the human in the loop.
  1. 1Prompt the coding agent to make the change you want.
  2. 2Never assume the generated code works just because it looks plausible.
  3. 3Run it and watch the code do the right thing with your own eyes.
  4. 4Only then deliver it — and include proof that it works.
Blog post — source

Try it — instruct the agent

Agent console

You're working on a small to-do web app. You want the agent to add a button that deletes a to-do item. Write the instruction you'd give it.

Agent behavior: Eager and literal — it does what you say, and fills gaps with assumptions when you're vague.

Check yourself

What is the core shift in working with an AI agent compared with a search engine?

The agent confidently says "Done!" What should you do?

Why do acceptance criteria and a scope boundary make delegation work better?

Your turn

Delegate one small, real task to a coding agent — but specify it well and verify the result instead of trusting it.

Try in Claude Code