Structure incoming requests

A proactive agent is an AI assistant that starts work by itself when something happens, such as a brief landing, an hour arriving or a fact changing. You don't have to prompt it each time. That makes it useful for recurring work, and it also makes it risky.

Dharmesh Shah, HubSpot's co-founder and CTO, argued that businesses treating agents as tools get wonderful tools, while those treating them as "future digital teammates" get something closer to transformation. Ethan Mollick of the Wharton School says working with AI is turning into something like management: hand over whole tasks, then "check the work later."

In practice, brief your intake agent the way you'd brief a part-time assistant on their first morning. Tell them what to watch for, which folders they can open, where finished work goes and who to ask when they're stuck. Then add the step most people skip, which is how you will check their work.

Suppose three bakery requests arrive overnight in your template shop. Turning each one into a structured brief by hand might take you 15 minutes. Checking a draft the agent has already structured might take 5. The agent doesn't save you the check. It saves you the typing.

A remit is one page that tells an agent what its job is and where that job stops. Give the agent a role name, such as Campaign Queue, and cover these parts:

• Trigger: what starts it. For example, a new request lands in the intake folder, or 7 a.m. arrives.
• Files and approved connections: the exact folders and apps it may open, listed by name.
• Where work lands: for example, a Briefs – Draft folder, never the production folder.
• Named owner and escalation: the person who owns this agent and who settles what it can't. In a shop of one, that's you.
• Agreed report: what it reports, when and where. For example: "After each run, post a summary in the queue log listing briefs drafted, fields left blank and the brand record version used."
• Pause and retire: how you stop it for a while, and how you end it for good without losing finished work.
• Closing check: the last line, saying what a person confirms before the run counts as done. For example: "Owner reads each draft brief against the original request."

If you can't fill in a line, you aren't ready to switch the agent on.

Every action the agent can take gets exactly one action setting:

• Always allowed: the agent does it without asking.
• Needs approval: the agent prepares it, then waits for you to say yes.
• Never: the agent doesn't do it at all. A person does it, outside the agent.

For intake, reading the queue and drafting a structured brief can be always allowed. Both stay inside your drafts and can be thrown away.

Sending a question to a client needs approval, because the agent would be speaking for your shop. Marking a brief ready for production also needs approval, because that hands the brief to the next agent.

Publishing anything, and deleting or overwriting a master (the template file that every buyer's copy is made from), are never allowed. Those two settings are the same for every agent you build in this course.

As this course describes them, Dots and Duet check with you before sending, paying, signing or deleting. Make your own settings at least that strict. Then test how your account actually behaves before you rely on it.

This course treats OpenAI Dots and Claude Duet as equals. In the course's description, each runs on its own cloud computer, connects to your apps and keeps working toward a goal. Pick the one that already works with your approved connections and that your helpers already use. Choosing the tool your assistant already knows is a perfectly good reason.

Next, decide how it runs:

• A single run does one job and stops. It suits tonight's three bakery requests.
• A continuing brief keeps the agent on duty. It checks the queue on a schedule or whenever a request lands. Because it keeps going, it needs a firmer agreed report and a pause you've actually tested.

You may have read that Canva AI 2.0 can run scheduled tasks in the background. Canva describes that as a research preview. The film behind this track found no recording of a Canva agent starting work by itself, so in this course you build the proactive start in Dots or Duet.

If you're unsure, start with a single run. Move to a continuing brief once the reports and the pause have worked a few times.

Every agent you build reads from the same brand record. That's one place holding:

• your brand rules as four lists: true, fixed, may flex, never ships without you
• your facts sheet (the count, sizes, what a buyer can change, claims, dates)
• first drafts of your brand consistency list (PDF), color-token sheet (ASE) and font-use guide

Date and version the record, for example "Brand record v3, 27 Oct." Each agent's report should then say which version it used. When a brief turns out wrong later, you can tell whether the agent misread the record or read an old one.

For the bakery template shop, "true" might say 30 editable social templates. "Fixed" might list the logo lockup, the color tokens and the two brand fonts. "May flex" might cover photo choice and background pattern. "Never ships without you" might cover prices, dates and any claim about what buyers get.

The intake agent reads this record. It doesn't edit it. Changing the record is your job.

A structured brief puts the same fields in the same places every time:

requester · audience · message · format and sizes · action · required brand elements · optional directions · deadline · approvals needed · source of each fact

Keep required brand elements separate from optional directions so later agents know which ones they can't change.

Say one overnight email reads: "Hi, can I get autumn posts like your Halloween set, with our pumpkin loaf, by Friday? — Rosa, Hilltop Bakery."

The agent fills in requester: Rosa, Hilltop Bakery (source: email 02:14), deadline: Friday (source: email) and message: autumn posts featuring pumpkin loaf (source: email). Brand elements come from brand record v3. Format and sizes stays blank, because Rosa never said, and approvals needed is filled in from your remit.

That blank is correct. An agent that writes "1080×1080, 1080×1920" because it sounds plausible has made up a fact. Run all three overnight requests through the same fields, then check each draft against its original. The next lesson covers what to do about blanks and duplicates.

Privacy is decided in the remit before the agent runs. You don't sort it out afterwards.

List the approved connections by name, for example "Intake folder (shared drive), Brand record v3, Briefs – Draft folder." Anything not on that list is closed to the agent. Keep customer lists, private campaign data and identifying images, such as photos of a client's staff or customers, out of any cloud tool you haven't approved. When the agent needs something sensitive, the remit points to the approved place where it's stored. Never paste the content into the agent's instructions.

Now build it. In Dots or Duet, create an agent named Campaign Queue. Paste in your remit and the three settings for each action, then attach the brand record. Choose a single run for your first test and feed it the three overnight requests.

Read its report. Does it name the record version, list the blanks it left and put everything in Briefs – Draft? If it does, the build passed the check. If not, fix the remit, not the output.