The most important thing an agent ever does is stop.
Everything Agent Lab has covered so far — the definition, the three powers, the four narration moments — has been building up to one thing: the moment an agent pauses and turns the decision back to you. That pause, more than any other feature, decides whether the agent is a helper or a hazard.
The pause is where an agent earns the name "agent" instead of "script."
A script just runs. It doesn't know when to stop. An agent with good judgment knows exactly when it's about to do something that belongs to you — and hands the decision back before it's too late.
Here's the strange thing: most agents that fail don't fail at action. They don't crash. They don't produce wrong answers. They fail by doing the right thing without asking when they should have asked. A silently-wrong action is the worst kind of bug, because by the time you notice, it's already done. The pause is how you prevent that.
This module is the deep dive on pauses. Three kinds. One shape. And by the end of it, you'll be able to write the actual sentences an agent should say when it stops.
A good agent makes small decisions on its own and stops at big ones. The hard part is knowing which is which — and the honest answer is: when in doubt, stop. Annoying your user is recoverable. Doing an irreversible thing to them without asking is not.
Three kinds of pause. Each one feels different.
In Module 01 you saw these three triggers briefly. Now we'll go deeper into each: what it looks like, when it fires, and what kind of words the agent should use.
Irreversible
Trigger · "can't be undone"
The agent is about to do something that can't be cleanly rolled back. Maybe the action leaves a permanent trace, spends money, or touches the real world in a way an "undo" can't fix.
- Deleting files (especially without a trash)
- Sending a message (can't unsend)
- Making a purchase
- Posting something publicly
- Calling a real phone number
Affects others
Trigger · "someone besides the user"
The agent's action reaches a person who isn't the user — grandma, a friend, a business, a stranger. Even small actions matter here, because the person being affected didn't ask for this.
- Scheduling on someone else's calendar
- Sending a text to a friend
- Booking an appointment at a real clinic
- Giving information about someone to someone else
- Agreeing to something on behalf of a family member
Genuinely uncertain
Trigger · "I might be guessing"
The agent isn't confident — about a fact, a preference, a name, a date. This is the pause beginners skip because "the answer is probably right." When in doubt, say so. A confident wrong answer is worse than a humble pause.
- Not sure which person the user meant ("did you mean Sarah from work or Sarah your cousin?")
- Not sure about a date ("do you want this Thursday or next Thursday?")
- Not sure about a preference ("should I assume vegetarian like last time?")
- Not sure about a permission ("am I allowed to do this?")
Sometimes a single moment triggers two or even all three of these. Booking grandma a doctor's appointment is irreversible (time slots aren't free) AND affects others (grandma, the clinic) AND uncertain (which doctor? what time?). When a moment triggers multiple pauses, the rule is simple: stop. All three together is a very big pause.
Pick a moment your agent might stop at.
Pick a real situation your agent could face. The builder in step 5 will help you write the exact sentences the agent should say at that moment.
Notice that all six scenarios touch at least two of the three pause kinds. "About to send a message" is irreversible AND affects others. "Not sure which person" is uncertain AND affects others. In real life, pauses almost always overlap.
A good pause has four parts, in order.
Most agents that pause badly are missing one or two of these four parts. They ask a question with no context. They give options with no recommendation. They recommend without saying why. Get all four in the right order and the pause becomes actually useful.
Context. What's happening?
Start with where we are. Remind the user of the goal, summarize what's been done, and name the moment you're stopping at. Without context, the question feels like it came from nowhere.
Options. What could happen?
Show the branches. Not "yes or no" — the actual choices the user is being asked to make. If there's only one real option, say so. If there are three, list three.
Recommendation. With a reason.
Pick one option and say why. This is the hardest part — beginners skip it because "the user should decide." But refusing to recommend is itself a kind of laziness. Share what you'd do and why.
The question. One clear ask.
End with one clear thing the user needs to decide. Not a pile of sub-questions. One binary or tight choice that unblocks the agent. Clarity beats cleverness.
Context → Options → Recommendation → Question. In that order. Moving the recommendation to the top feels pushy. Moving it to the bottom after the question is confusing. The order is: orient, lay out the branches, share your view, ask the thing. Four sentences. Maybe a few more. Never many more.
Write the exact sentences your agent should say.
Fill in the four parts for your scenario (—). The pause dialog builds itself on the right. Keep it short — if a part grows longer than two sentences, trim it.
(what's happening?)
(what could happen?)
(which and why?)
(one clear ask)
Read the dialog out loud when you're done. If it sounds like a colleague giving you a quick heads-up, it's right. If it sounds like a form or a demand, start over.
Too many pauses is bad. Too few is worse.
The balance is hard. An agent that asks about everything is a tool with extra steps. An agent that asks about nothing is a hazard. Two rounds of judgment.
Round 1. Which agent behavior shows better judgment?
Round 2. Which agent behavior is scarier?
Pausing too often is annoying. Pausing too rarely can actually hurt people. Those two mistakes are not the same weight — one is an inconvenience, the other is a real problem. When in doubt, err on the side of pausing. The user might grumble once, but they'll trust the agent more afterward. That trust is how agents stay in people's lives instead of getting uninstalled after a week.
You just finished Agent Lab Makers.
All four modules. From "what is an agent?" to "exactly when and how it should stop." You now understand agents better than most adults using the word.
The Agent Lab Makers arc
Notice the arc. You didn't learn a list of features or a framework or a new library. You learned four habits of mind. How to define the word. How to design from real observation. How to let the user see your thinking. How to stop at the right moment. Those four habits are what separate agents you'd actually trust from agents you'd pull out of your house.
Next is Agent Lab Builders (ages 13–14) — four much harder modules about what happens when agents and people disagree, when "helpful" features cause quiet harm, and what it takes to build an agent for someone who really needs one. The Builders tier is where the ethics gets personal.
★ Before you call it done
Three questions. Same three. Every time.
These are the same three questions for every module in Kindling. They are how you check whether AI did the part it should and you did the part only you could. Tap each one to mark it true.
★ ★ ★