Five questions to ask before granting autonomy
What can the agent do? Tool permissions explains the effective catalogue. Why did information return? Governed memory follows provenance and access. What happens after failure? Resilience examines recovery. Observability connects evidence and costs; account administration identifies who controls the settings.
Control before, during and after work
Before execution, select the agent, sources, tools and approval requirements. During execution, inspect the plan, answer questions and stop unsuitable work. Afterwards, verify the deliverable, receipts, usage and incidents against the original request.
Grant autonomy by capability and context
Reading a source, editing a document and sending mail have different effects. Configure access accordingly and keep consequential decisions explicit. Actual protection depends on the instance, tools, models and operating procedures; the presence of controls is not a certification.
Controls visible in daily work
Email approval can cover the exact message stored before sending. Document reading, editing and sharing remain separate permissions. An uncertain task effect is not treated as success by default.
Observe these controls in a trial: edit a shared document, reject a delivery, recover a file transfer or inspect a cost without known pricing. Email approval provides a complete example.
Personal filing does not share a document. An agent’s request to open it is not proof of reading. A stop request is not proof of remote stopping. The control pages explain these distinctions so each stage has an appropriate action.
You also choose application access
An application inside a document does not inherit its author’s rights. It uses the reader’s rights plus consent for each Dataset binding and content version. Sharing the page, sharing data and authorising the application remain separate. Understand these controls.