Working with agents

Choose the right way to work with your agent

Otto gives you three ways to ask an agent to help: a one-off task, an ongoing responsibility, or an automation. Start with the smallest option that fits the work. That makes it easier to review the result, set the right access, and avoid repeating a workflow before it is ready.

Compare the three options

Choose	Use it when	Example
One-off task	You need one result now, or you are still testing an idea.	“Compare these three vendors and recommend one.”
Ongoing responsibility	An agent should keep helping with a role that needs your context and judgment over time.	“Help me plan this launch and keep track of the next decisions.”
Automation	A proven workflow should run on a schedule or at a defined time without you starting it each time.	“Every Monday, summarize last week’s support themes.”

These are choices about how work is run, not permission levels. An agent still needs the relevant connected app access, and actions that affect an outside service may still require approval.

Start with a one-off task

Use a normal conversation when the outcome is new, unusual, or easy to review. This is the best way to learn whether the agent understands the request and whether it needs a file, a connected app, or an approval.

Be specific about:

the result you want;
the information the agent may use;
what it should not change or send; and
what it should bring back to you if it cannot finish.

For example, ask for a comparison and a recommendation, but say whether the agent should only draft the recommendation or is allowed to contact a vendor.

Keep a responsibility with an agent

An ongoing responsibility is useful when the same role repeatedly benefits from shared context and follow-through. Give it a clear boundary:

State what the agent owns and what a successful result looks like.
Identify the sources it may use and the apps it may access.
Say what it may do independently and what must come back to you for a decision.
Decide where updates should live so work does not get scattered across conversations.

A responsibility is not a license to act broadly. Connected accounts are shared by every agent on the Otto, so keep the device's connections and the approval boundary as narrow as the work allows. If two roles need the same context, access, and judgment, keeping them with one agent is usually simpler than creating another.

Turn a proven workflow into an automation

Create an automation only after you have run the task manually and can describe its definition of done. Before saving it, make sure you know:

Trigger: when it should run, such as a schedule or one-time run.
Owner: which agent should perform it.
Request: the exact input, output, and stopping or escalation rule.
Access: whether it may use connected apps at all, read only, or read and change. One level covers every connected app.
Controls: approval settings and usage limits.
Delivery: where the result should appear and whether Otto should notify you.

Open Automations, choose to create one, then review those settings before selecting Save. Use Run now to validate the first run. Keep the first few runs easy to inspect; adjust the request or delivery setting before relying on the automation.

Automations are good for recurring checks, reports, and other repeatable work. They should not silently replace a decision that still needs your judgment. If a run fails, inspect its history and fix the smallest specific problem—such as a disconnected account, unclear request, or insufficient access—before running it again. Do not recreate the automation just because one run failed.

A quick decision guide

Need an answer or draft once? Start a one-off task.
Need a role to retain context and help you make progress? Give an agent a responsibility.
Need the same check or report at a predictable time? Create an automation, after testing it manually.
Need the agent to send, publish, purchase, or change records? Review the approval and app-access settings before enabling that action.