Staying In Control
Autonomy is worth handing over only when you can take it back. Every other control rests on one brake the agent cannot refuse. Your input lands at the next safe point between steps, so a redirect never breaks the run.
I hand the agent real autonomy for one reason: I can stop it any second. That brake sets the size of the job I am willing to leave alone.
Control model
Put the agent behind the wheel of your car. Two things make that safe.
It asks before every sharp turn, so the risky ones wait for your yes. You reach in at any second without wrecking the car.
Take the asking away and you watch every step. Without a way to reach in mid-drive, you trust the car with your eyes shut. Hold both and you can hand over the keys.
Taking the wheel is easy. The hard part is landing your words without wrecking the step already running. The agent is usually mid-tool when you interrupt. So the moment your input lands matters as much as what it says. Run all three scenarios below and watch where your message goes.
01Control is the price of autonomy
Start with the brake, because every other control assumes it works. It takes effect now, and the agent holds no way to wave it off. The more the agent does on its own, the more that brake carries. A stop lands at the next safe point, so the spacing between points sets your stopping distance. Push the safe points apart below and watch the brake slow down. Remove them and the brake never lands.
02Where your input lands is the hard part
When you reach in, the agent is usually partway through a tool call. Your words cannot land there. A tool request and its result stay paired, a rule from the first idea in this guide. Slip another message between them and the conversation breaks.
So the runtime queues your words. It holds them until the current step ends, then lets them land at that safe point. You see an agent that redirected the moment you asked. Underneath, it spent one beat holding the pair intact.
The idea
By steering I mean this: your words take effect at the next safe point. You never see that wait when it works. That wait is the reason a new instruction can arrive mid-run at all.
Stop it
Halt now. It winds down to the next safe point and stops cleanly, with the work so far kept.
Steer it
Keep going, but differently. You add a message and it changes course at the next safe point, no restart.
Walk away
Let it run while you aren't watching. Come back to the result later. Out of your sight, not out of your control.
03Three ways to take the wheel
Control splits into three moves. An agent worth trusting gives you all three.
- Stop it. The brake lands at the next safe point, and everything finished before it survives.
- Steer it. You type while it runs, and the agent redirects at the next safe point without starting over.
- Walk away and come back. The run keeps working in the background while you do something else. Backgrounding a whole job is the subject of the subagent piece.
Each move answers the same need: do something other than sit and watch. Two of them wait for a safe point. All three keep the work you already paid for.
04Permission is the other half of control
Permission and steering are the same control from two directions. The agent stops and asks before a risky action, and the permission piece walks that path end to end. Steering is you reaching in without waiting for the question. Permission covers a run you left alone, and steering covers the one in front of you.
The cost
This control costs you a beat. A stop waits for the current step to finish. A redirect can land after the agent already did the thing you wanted stopped.
The bill comes due when you are away from the screen. A missing question or a confident wrong turn runs unchecked there. Watch one measure: the wait between reaching for the wheel and your words landing. Keep the safe points close together, and that wait stays short.
Decision
I judge an agent mid-run, with two questions. Can you stop it, redirect it, and walk away without breaking it? Can you change your mind and have that change land cleanly? One no on either question means you watch that agent, so keep the work small. The next piece takes what keeps a run alive when a step breaks on its own.