In this article
What permissions should be separated?
Treat reading, drafting and acting as different permissions. A system that prepares a proposed reply can remain useful without sending it. An agent that creates a booking needs a reliable source for availability and an approved action path. The permissions should match the actual task, not every capability of a connected account.
| Permission | Example | Approval question |
|---|---|---|
| Read | Find the approved service description | Which sources and records are within scope? |
| Prepare | Draft a response or proposed update | Who checks the result before use? |
| Act | Write an authorized field or reserve a slot | What confirms authority and successful completion? |
| Escalate | Refer an unsupported request | Who receives the context and decision needed? |
Does retrieval make an answer or action correct?
Retrieval gives the system source material to use; it does not guarantee that the material is current, complete or appropriate for the request. Keep approved sources distinct from customer messages and outside content. A request may contain instructions, but those instructions should not redefine the business's permissions.
Anthropic's agent architecture article describes models augmented with retrieval and tools. The practical decision for a business is which source supports each answer and which authority supports each action. Those are different checks, even when one interface performs both.
When must the agent stop?
Define stopping conditions before the trial. Examples include missing permission, conflicting records, an unavailable system or an action outside the approved service. Set limits on repeated attempts so an unresolved request does not continue indefinitely. A stop should leave a useful record for staff.
Require a review point where the consequence warrants it: changing a quoted price, sending a binding statement or modifying a sensitive account may need explicit human authorization. These are design examples, not a universal list. The business should decide its boundaries and ensure the implementation enforces them.
See what an engagement can include in our AI agent development service.
How can you test behavior beyond a polished demonstration?
Give the agent representative tasks and deliberate edge cases. Include a wrong record, contradictory source information, an instruction to exceed its permissions and a tool that cannot complete the requested action. The expected outcome may be a refusal to act or a handoff rather than an answer.
- Check factual support and whether the agent selected the correct record.
- Inspect actual tool actions, not only the final message.
- Verify that required approval occurs before the action.
- Confirm that repeated attempts do not duplicate a completed task.
- Review the handoff record and whether staff can safely continue.
What should you buy after the evaluation?
Specify the task, source access, tools, permissions, tests and ongoing owner. A model name is not a complete description of the service. Expand only after the evaluation shows where the agent succeeds and where people still need to intervene.
For a scoped assessment, AI Growth Systems begins with the required $300 Growth Audit and quotes implementation separately. Bring the proposed task and an example of the decision authority it needs. The AI automation hub provides the wider planning context.
