An agent is an assistant that runs a multi-step task without you prompting each step. What that looks like, which tasks it suits, how to give one a job it won't mess up, and why the boring word 'budget' matters more than the exciting word 'autonomous'.
Everything so far has been one step at a time: you ask, it answers, you ask again. An agent removes the "you ask again". You give it a goal and some tools, and it plans a step, does it, looks at the result, and decides the next step, until the job is done or it gets stuck. Lesson 6's workflow, run by the assistant rather than by you.
This is the part of AI that gets the most breathless coverage, so let's be plain about it.
What an agent is
Plan, act, look, repeat. You're at the start, at the end, and at every irreversible step in between.
Three things make something an agent rather than a chat:
It can use tools. Search the web, read and write files, run code, call the connectors from lesson 8, click through a browser.
It loops. It looks at what a step produced and chooses the next step itself.
It stops. When done, when stuck, when it hits a limit you set, or when the next step is something you said to ask about first.
You've probably used one without the label. "Deep research" modes that search dozens of pages and write a report are agents. Coding tools that change many files and run the tests are agents. Assistants that can operate a browser to fill in a form are agents. Some products call themselves agents and are a chat with a nicer name. The test is the loop.
What agents are good for today
Honest list, from most to least reliable:
Research with a clear question. "Find the five largest suppliers of X in the UK, their approximate size, and whether they publish a sustainability report. Cite sources." Many searches, one report, you check the citations.
Multi-file work on your own material. "Go through these forty meeting notes and build a table of every commitment we made to clients, with dates." Tedious for you, mechanical for it.
Software tasks. Small, well-described changes to code, with tests. This is where agents are most mature, which is why lesson 10 uses one.
Repetitive admin with a human at the end. "Draft replies to every email in this folder using my skill; leave them as drafts." It does the forty drafts; you send.
Where they're weak: open-ended goals ("improve our marketing"), tasks where the right answer requires judgement you haven't written down, and anything where a wrong step is expensive and hard to reverse. An agent is a diligent junior with no context beyond what you gave it. Brief it like one.
Giving an agent a job it won't mess up
Five parts, and they'll feel familiar:
The goal, with a definition of done. Not "research competitors" but "a table of our six named competitors with price, headcount and last funding round, each with a source link. Done when all six rows are filled or marked 'not found'."
The material. Files, links, the skill with your format. It only sees the window and what its tools fetch.
The limits. How many steps or how much time. Which tools it may use. What it must ask before doing. "Read anything, write only to this folder, never send email, ask before deleting." Good agents let you set these as settings; if not, say them in the brief.
The checkpoints. Where you want to see progress before it continues. For a long task: "Show me the first three rows before doing the rest." Cheap, and it catches a wrong interpretation before it's multiplied by forty.
The budget. Agents consume a lot more than a chat does, because every step re-reads the window, and the window grows with every result. A runaway agent can burn a day's allowance in an hour. Set a limit, and prefer tasks with a natural end.
Reviewing the result
Lesson 3 applies, with one addition. Beyond checking the specifics, look at the trail: most agents show the steps they took, the searches they ran, the files they touched. Skim it. You're looking for a step where it guessed, got an error and pressed on, or wandered off the brief. That trail is also how you improve the next brief: wherever it wandered, you under-specified.
For anything that wrote, sent, changed or deleted, review before it's final, not after. The whole value of a checkpoint is that it's before.
Where this is going
Agents are getting better fast, and the direction is clear: more of lesson 6's workflows will run themselves, with you reviewing rather than driving. The people who'll be good at that are the ones who built the workflows by hand first, because they know what each step should produce and where it goes wrong. That's you, if you did lessons 6 and 7. The skill isn't operating agents. It's writing a brief with a definition of done, limits and checkpoints, and reading the result like a manager rather than a spectator.
The habit
Bounded goal, definition of done, limits, checkpoints, budget. Review the trail, not just the output. Reversible tasks first.
Try this today
Pick a research question from your work with a clear definition of done, like the competitor table above. Give it to a deep-research mode or an agent. Set a checkpoint: "show me the first two rows first." Check the citations on the final result. Count how many were right. That number is your calibration for how much to trust the next one.
Check yourself
5 questions. Everything is in the lesson above. Get them all right and the lesson counts as done.
1. What distinguishes an agent from a chat assistant?
The loop is the agent. Tools let it act; stopping rules make it safe.
2. Which task is best suited to an agent today?
Bounded, mechanical, on your own material, with a definition of done. The others need judgement you haven't written down.
3. Why does "budget" matter so much with agents?
Set a limit on steps or time, and prefer tasks with a natural end.
4. What's the purpose of a checkpoint like "show me the first three rows before continuing"?
A wrong reading of the brief is cheap to fix at row three and expensive at row forty.
5. When reviewing an agent's result, what should you look at besides the output?
The trail shows where it wandered, which is where your brief was under-specified. That's how the next brief gets better.