Agentic software works within boundaries

Useful because it is bounded

The question people ask most about AI agents is not "what can it do?" It is "what will stop it doing the wrong thing?" In agentic software, the answer is designed in, not added afterwards.

Every agent works within boundaries the organisation sets. Those boundaries are what make it safe to hand over real work.

The four boundaries

1. Permissions. Each agent has a defined role. It can only use the information and actions that role allows, through the same permission checks and business rules as your staff.
2. Approvals. Work that matters, or is hard to undo, waits for a person to review and approve before it goes ahead.
3. Escalation. When an agent reaches something outside its role, such as an unusual case, missing information or a judgement call, it should stop and hand it to the right person with its working attached.
4. Intervention. People can see what agents are doing, stop further work, and review or correct what was done. Delegation is never surrender.

Permissions and approvals are enforced by the software itself. Escalation depends on the agent recognising when something falls outside its role, which is why it is tested before an agent is trusted with real work.

The goal is not an agent that can do anything. It is one you can trust with something.

Setting the boundaries to the stakes

Not every task needs the same limits. A useful rule of thumb is to set them by the stakes and by how easily a mistake can be undone. Drafting a routine reply can run with light review. Changing a contract or a payment should wait for a person.

The boundaries also leave a record. What was done, what was approved and who approved it are all visible afterwards, which is what accountability needs.

For a worked example of a decision carried through within limits, read Clicking a button is rarely the real outcome.