Your First Agent in a Weekend
One task. Two days. That's the entire scope.
A chatbot answers when you ask. An agent works when you don't. The gap between those two is smaller than the industry makes it sound, and you can cross it in a weekend without writing a line of code. What you can't do is skip the last step, so read all the way to Sunday afternoon before you start.
Friday night: pick the task.
Run every candidate through three questions. Could you explain the task to a new hire in five minutes? Would you trust a smart intern to do it? Can you check the output against something that doesn't lie, a source document, a total that has to reconcile, a known benchmark?
Three yeses and you have your task. Good first candidates: a morning summary of yesterday's numbers against target, the weekly client status draft, a scan of new invoices for anything that doesn't match your pricing history. Bad first candidates: anything that moves money or talks to a customer, and anything you can't verify.
One more filter. Pick something you already do every week and mildly resent. Resentment is a fine compass; it points at repetitive work with clear standards, which is exactly what agents eat.
Saturday: teach it.
Open whatever chat tool you already use and work the task the way you'd train that new hire. Give it the context, the source data, and the format you want back. Correct what it gets wrong. Run it again. Correct again. You're done teaching when the output comes back right five times in a row, not once. Once is luck.
Then do the step most people skip. Stop chatting and write the winning instructions down as a standing document: the role, the inputs, the steps in order, the exact output format, two examples of good output, and the definitions it kept fumbling. That document is the asset you built this weekend. The chat window was scaffolding.
Sunday: schedule it, then alarm it.
Give the task a trigger so it runs without you. Any mechanism works: the scheduled-task feature inside your AI tool, an automation platform, a calendar-driven workflow. The vendor matters far less than the two requirements that come next, because no agent should run unattended without both.
First, a verification step. Before the output reaches you or anyone else, it gets checked against something objective. Does the report contain today's data? Does the total reconcile with the source? Are all required sections present? Output that fails the check does not ship.
Second, a fail-loud alert. If the agent can't do the job, missing data, broken access, empty source, you get a message that says so, by whatever channel you actually read. Treat silence as a failure state. Never as a success.
Skip those two and you will eventually learn the expensive version of this lesson: an agent with missing data doesn't stop. It fills in the template anyway. It will produce something shaped exactly like the report you asked for, on schedule, indefinitely, with flawless formatting wrapped around nothing. An unwatched agent never confesses. Build the alarm before you grant the autonomy.
| When | What | Done means |
|---|---|---|
| Friday night | Pick one task using the three questions | Explainable, intern-safe, verifiable |
| Saturday morning | Teach the task in chat | Right output five times running |
| Saturday afternoon | Write the standing instructions | Someone else could run it from the document |
| Sunday morning | Put it on a schedule | Runs without you touching it |
| Sunday afternoon | Add the check and the alarm | Wrong or missing output reaches you as an alert, not a surprise |
What Monday looks like.
For the first two weeks, read every output before you rely on it. The review takes minutes, and it changes your role: you're the editor now, not the producer. When corrections come up, and they will, add them to the standing instructions instead of fixing them in chat. The document gets smarter. The agent inherits it every run.
The second agent goes faster than the first. The pattern is identical every time: pick, teach, write, schedule, verify. By the fifth one you'll notice the real shift, which was never about any single task. You've started converting how your business runs from memory into writing, and written operations compound.
Some owners run this loop themselves and enjoy every step. Others want the first ten built, verified, and humming while they stay on the work only they can do. If you're the second kind, that's a Discovery call.
New pieces land here first
One email when something publishes. No drip sequence, no reselling your address.