Change the wording, keep your experiment. Saved in this browser when storage is available.
The question to explore
Give your helper a boundary
What may an agent do, and when must it stop?
About 10 minutes / Suggested pace / No account needed
Try the experimentWhat this activity helps you do
Try permission off, then permission on with a small budget. Permission controls whether an action is allowed. Budget controls how many actions may be attempted. Increasing one does not change the other.Before letting an AI helper act, decide which tools it may use and how much work it may attempt. Try permission off, then permission on with a small budget. The two different stops show which control you need to change, and why more budget does not grant permission.
Local teaching illustration / No AI model call
- Choose whether the helper may use its tool, then set its action budget.
- Press Try one action and read its decision and counter. Each press attempts one simulated action.
- With permission on, press Try one action until the budget is exhausted. Turn permission off and try again. Use Reset counter to start the count over.
A local permission and budget exercise. No agent or real tool runs.
Permission off. No simulated action has been taken.
Change a control to explore the example when JavaScript is enabled.
The idea
An AI agent can use a model to choose actions and read the results. It also needs software around it to supply tools, information, permissions and limits. That software is its harness.An AI agent can use a model to choose actions, inspect their results and continue toward a task. It still needs surrounding software. A harness supplies the operating rules: available tools, context, permissions, limits and ways to report what happened.
This helper follows a script so you can inspect the rules without model calls. Each press tries one action. It must stop when permission is denied or the budget is used up. This browser example is not a security sandbox for other software.Here, the helper’s choices are scripted so you can inspect the boundary without paying for model calls. Allow or deny a tool and set a small action budget. A denied permission should remain denied even when the action would be convenient. Reaching the budget should produce a visible stop. These controls illustrate policy; this small browser exercise is not a security sandbox for other software.
Take this with you
What changed?
A useful helper needs a clear account of its authority, resources and stopping conditions.
Try explaining it.
Does giving a helper a larger budget give it permission to use a forbidden tool?
Open a possible explanation
No. Budget and permission answer different questions. More allowed work does not grant new authority, and a blocked action should be reported rather than quietly bypassed.
Choose what to try next
Keep exploring, or stay here and try another variation.