Hey All, working on my agent academy work, here’s some initial guidance that may help on the next step. Talk to you soon. Let’s get into it.
Morning all.
Last week I wrote about getting an agent unstuck by remembering what we were actually trying to accomplish. I stepped in, did the design review, and handed the work back.
This week, let’s make the starting point a little less ambitious.
If you’re curious about agents, you don’t need to begin with a team of them building your next business. Pick something small enough that you can tell whether the result is any good.
Boring is fine. Boring work still needs doing.
Start with something you understand
A good first assignment is turning meeting notes into a follow-up draft.
You know what happened in the meeting. You know who agreed to what. You can check whether the draft is accurate before anybody receives it.
That gives you room to learn without handing an unfamiliar system your entire operation and hoping for the best.
For this exercise, we’ll use made-up notes. No customer information, private discussions, or connecting your inbox required.
Sample meeting notes
We discussed a customer onboarding guide.
Alex will draft the outline by Tuesday.
Priya will check whether the existing screenshots are still accurate. No deadline was agreed.
We discussed a Friday launch, but did not decide.
The next meeting has not been scheduled.
You could paste those in and say, “Write a follow-up email.”
You might get a perfectly useful answer. You might also get a confident announcement that the guide launches Friday.
Apparently we’ve made a decision. Lovely.
Give it a job you can check
Try this assignment instead:
Turn the sample meeting notes below into a follow-up email draft for the people who attended.
Use only the supplied notes. Do not invent decisions, deadlines, attendees, or commitments.
Include three short sections: agreed actions, open questions, and next meeting.
Preserve each named owner and agreed deadline. If a deadline or meeting date is missing, say it has not been agreed. Keep proposals separate from decisions.
Keep the draft under 200 words.
Before writing, identify any missing information that prevents an accurate draft. If the missing information can be clearly marked as undecided, proceed.
Return the draft and a short list of anything I need to confirm. Stop there. Do not send the email, schedule a meeting, or update any other system.
[Paste the sample notes here.]
There’s a result, an audience, a source, a quality standard, and a stopping point.
Those details give you something concrete to review. They also make the difference between preparing a draft and taking action on your behalf explicit.
For this first exercise, leave the sending tools disconnected. Written instructions are useful, but you should also control what the system can actually access.
Now read it like the person responsible
Don’t just ask whether it sounds professional.
Compare it with the notes.
Did Alex keep the Tuesday deadline? Did Priya acquire a deadline nobody agreed to? Is Friday clearly a proposal? Did the agent quietly schedule the next meeting in a sentence?
Here’s a deliberately flawed example:
Alex will finish the outline Tuesday. Priya will update the screenshots by Thursday so we can launch Friday.
It reads smoothly. It also changes “check” to “update,” invents Thursday, and turns a possible launch into a commitment.
That’s three problems in one sentence.
A useful correction would be:
Priya agreed to check the screenshots, not update them. No deadline was agreed. Friday was discussed but not approved. Correct those points and check the rest of the draft for invented commitments.
Then read the revised version. Asking it to check its work doesn’t remove your need to check it.
Keep the first lesson small
This exercise works in an ordinary AI chat. By itself, it doesn’t demonstrate an autonomous agent running a workflow.
It lets you practice something you’ll need when you do give an agent more room: explaining the assignment and judging the result.
Once that feels comfortable, you can decide whether a connected tool would actually help. Maybe preparing the draft in your email application would save a step. Sending it is another decision.
You don’t have to grant every capability at once.
That’s the approach I’m taking with Agent Academy: start with work you understand, try something concrete, and learn from what comes back.
Episode 1 is available here if you want to start with what an agent actually is.
For now, a short draft with the right commitments is a perfectly respectable result.
If you want help making AI useful in your team’s actual work, work with me at Startups and Code.