Choose approvals around the actions you intend.

Describe what is allowed to happen

Start with the result and the operations needed to reach it. “Prepare a supplier recommendation” can mean reading files and drafting a brief. It does not necessarily mean messaging suppliers or approving a purchase. Write the boundary in ordinary language before choosing a mode so the setting supports an understood task.

Separate reading, creating a draft, modifying an existing resource, and making a change visible to someone else. The same Tool may support more than one kind of action. Choosing a mode without reviewing the enabled capabilities can leave you with a broader scope than the brief suggests.

Use the meanings documented in Workspace

Session settings include Ask, Notify, and AUTO. Ask is the documented choice when you want to review changes first. AUTO approves all enabled actions. This makes the Agent’s selected capabilities relevant to the decision: an Agent that can only prepare work and one that can operate external systems present different review needs.

Read the current description of Notify in the product before relying on it. This guide does not infer that notification means an action waits for you, or that every operation uses an identical gate. When a specific action requires your review, use Ask and confirm how that action appears in the task’s actual flow.

Make the review point concrete

Consider a fictional supplier-review task for Cedar Studio. The initial request is to compare two proposals and save a recommendation. Later, the owner may decide to request revised terms. That later decision is a new action with a different audience and evidence requirement.

The table identifies what a reviewer needs to see before expanding the scope. It is a task-planning example, not an inventory of product-enforced permission categories. Write the actual restriction in the task and configure the Agent accordingly. Do not assume that a draft’s presence authorizes its delivery.

Review the Agent before changing the mode

Open the Agent’s configuration and inspect its instructions, knowledge, Skills, and selected Tools. Remove capabilities that the task does not need when the product allows that choice. Confirm which account, Project, or environment supplies the context. An instruction to be careful is not a substitute for knowing what the Agent can operate.

Begin with a small, representative task in Ask. Review the proposed change and inspect the result afterward. This tests both the instructions and the path the action takes. If the Agent asks for an unexpected capability, identify the task step that requires it before expanding access.

Review recurring work separately

A Session approval choice and unattended computer work are different controls. For a Loop or Schedule to run scripts and commands while your browser is closed, Allow computer jobs must be enabled and its Agent must permit unattended work. Answer only, Research, and Always ask Agents cannot run unattended computer jobs.

Test one run before configuring recurrence. Decide where results should be saved and what should happen when a source is missing or an action cannot proceed. Hosted work may run in the cloud, while a task using a connected computer still needs that machine online. AUTO alone does not establish these execution conditions.

Resolve uncertain results before another attempt

If a task stops after beginning a change, inspect the relevant result before repeating it. A saved file, modified document, submitted form, or external message may have a different completion state from the conversation. Ask what is confirmed and what remains uncertain; avoid treating silence as proof that nothing happened.

For a draft-only task, reopening the file may settle the question. For an external action, check the destination or available history before authorizing another attempt. Keep the next instruction narrow: reconcile the specific result, then continue only the unfinished work. This prevents an approval decision from accidentally becoming permission to repeat a completed operation.

Try it with a small example.

Before you use the result.

• The task names which external actions are outside scope.
• The Agent’s enabled capabilities are understood.
• The review point is tested with a representative change.
• Unattended work is configured and checked separately.
• Uncertain results are inspected before another attempt.