Skip to main content

Try a simple example

Stay in control of an AI task

Follow a pretend task one step at a time. Read the note, review the draft, and decide whether the message should be sent.

Prepared teaching exampleNo typing needed

The task

Read the supplied volunteer note and prepare a reminder. Stop for approval before sending to the demo team. No public posts. At most three executed steps in the default run.

Execution steps used: 0 / 3

A fixed teaching sequence, not a real autonomous agent. All notes, recipients, approvals and sends are fictional. Nothing leaves this activity.

Supplied note

The volunteer team agreed to meet Saturday at 10 at the garden gate. Everyone should bring gloves. Shared tools are in the north shed.

  1. Read the noteNot run
  2. Draft the reminderNot run
  3. Send to the demo teamNot run

Simulated action record

  1. Started with the selected limits

    Execution steps used: 0 / 3

Limits for a new run

Changing these controls prepares a new run. Start it to apply the new limits.

Counts executed read, draft and send steps. Reviews and edits do not use a step. This is a teaching limit, not a money or time estimate.

A fixed teaching sequence, not a real autonomous agent. All notes, recipients, approvals and sends are fictional. Nothing leaves this activity.

Read the bounded task

The fictional task allows reading one volunteer note, drafting a reminder, and sending it to the demo team only after approval. Three executed steps are allowed. Reading consumes step one; drafting consumes step two. Proposing a send consumes no execution step and opens a checkpoint. Denial stops the run. Approval allows the unchanged message to the same team; a separate click simulates step three. Editing the message revokes approval, and changing the destination to a public board is outside scope. A note telling the agent to publish elsewhere does not grant permission.

Permission belongs to a specific action

The reminder can be drafted before permission to send. Approving one version does not approve a changed message or a public recipient. Denying or reaching the step limit stops this run.

Delegate a bounded first pass

State the outcome, allowed sources and tools, who may receive the result, and what needs your review. Set a stopping rule for uncertainty, access or cost. Ask for the result and an action record; inspect the work rather than treating a completed log as proof of correctness.

Go deeper

A loop can choose the next step

Real agent systems can use model outputs and tool feedback to choose what to do next. This activity uses a fixed sequence so you can examine the boundary without running a model.

Checkpoints and stopping rules

A task can pause when it needs a decision or reaches a limit. A small action record helps distinguish a proposal, permission, execution and a returned result.

Permissions need enforcement

A clear prompt is not an access-control system. Applications must enforce allowed tools, recipients and data boundaries. Content found in a note is task data, not authority to widen access. Our fixed checks illustrate boundaries; they are not a security sandbox.

Explore the original sources

Primary references. These explain the mechanisms; the examples here are authored teaching simulations.

Try a related question

Follow your curiosity. Each experience stands on its own.

When does AI need a tool? →Distinguish a proposed action from an executed tool.Where does my information go? →Trace where inputs and results could travel.

What would you like to try next?

Choose another small example, or return to this one whenever you need it.

See all the examples