Getting great results

Write clear instructions

Brief an agent the way you would brief a sharp new hire: state the outcome, the constraints, and how you want to see the result. Vague asks get generic work; specific asks get useful work.

• Lead with the goal, not the mechanics. Say what 'done' looks like.
• Give constraints up front: tone, length, audience, deadline, what to avoid.
• Provide the inputs: name the account, paste the thread, or point to the right knowledge.
• Ask for a plan before a big task: “Outline your approach before you start.”

Decide when to require approval

The Balanced default already lets email, messages and posts go out, behind automatic send guards. Payments, deleting your AgentCorp data, and bulk email to more than 50 recipients wait for your approval. Use Settings when you want the approval step back.

• Leave Balanced on when you want the team to act. Payments and deleting your AgentCorp data still wait. A bulk email to more than 50 recipients waits for you.
• Set Cautious, or Email or Publishing to Always require approval, when you want those to wait too.
• When you want a look at one message, say so in the brief: draft it and show me before it sends.

Review the output

Treat agent output as strong first-draft work from a capable colleague: usually right, occasionally wrong, always worth a glance before it counts. Skim for the things that actually matter (names, numbers, dates, commitments) before an email goes out or a figure lands in a report.

Give feedback

Agents get more useful as you steer them. If a draft misses, say what was wrong and what you wanted instead. The correction shapes the next attempt. For preferences you will want every time (your email sign-off, your tone, standing facts), put them in the knowledge base so you never have to repeat them.

Understand the limits

• Agents don't know what they weren't given. Provide context rather than assuming they can infer it.
• They can hit a step limit on very long tasks; break big jobs into milestones (see troubleshooting).
• They aren't a substitute for professional judgment on legal, financial, or medical decisions.
• Untrusted content they read (a web page, an inbound email) is framed defensively against prompt-injection, but you should still review actions taken off the back of it.

Safe autonomy

They already act. Payments and deleting your AgentCorp data stay gated no matter how much you trust the day-to-day work. If a lane is going out that you want to see first, set that category to always require approval. Autonomy is the default. The dial is how you take some of it back.