Everyone is talking about agents. Almost nobody explains them for people who run a business, not a codebase. This is that version.
You already use the first two. The third is the one that does the work on its own. Same family, very different jobs.
Answers when asked, then waits for you. Smart, but it just sits there.
When X happens, do Y. Reliable, but it can't decide. You wired every move.
You give it a goal. It picks the steps, does the work, checks itself, and finishes.
Strip the jargon and every agent is the same three beats, reason → act → observe, repeating toward your goal until a "done?" check says stop.
Think of a great assistant you do not micromanage. You hand them the goal. They figure out the next step, do it, check their own work, and go again, only coming back when it is actually done, or stuck. An AI agent is that, running on its own.
Each lap is one step. The agent keeps circling (act, observe, act, observe) until the "done?" check passes, or a guardrail (a hard limit, like a max number of tries) stops it. The Observe beat is what makes it work: it reads its own result instead of assuming it worked. That is exactly why the "done?" check is everything.
An agent is worth the setup only when a couple of things are true. Run your task down this ladder before you build anything.
Small business owners say the same admin work eats their week: chasing payments, chasing approvals, and finding new clients. Here is each one as an agent. You give a goal, it loops, you get a result.
You give
"Email every client whose invoice is more than 7 days overdue a polite reminder. Don't double-send."
It loops
Reads your invoice list, finds who is overdue and not yet reminded, drafts each note, checks the list again.
You get
Every overdue client reminded, a list of who got what, and nobody nudged twice. Chase payments, handled.
You give
"Find who still owes me an approval or a document, and send a friendly nudge with what's outstanding."
It loops
Scans your tracker for items waiting on someone, drafts a specific nudge each, confirms none already cleared.
You get
Every stalled approval followed up, in your words. Chasing approvals, off your plate.
You give
"Find local businesses whose website talks all about 'we' and never the customer. Rank the best rebrand prospects."
It loops
Searches local businesses, opens each site, scores the "we vs you" copy, keeps only the strong matches.
You get
A ranked list of ready-to-pitch clients, with why each made it. Finding clients, while you sleep.
Same shape every time: a goal, a loop, a checkable result. Swap in whatever eats your week.
It will happily produce confident, polished, wrong work and call it finished, unless you tell it exactly what "done" means and give it a way to check.
Lock your finish line: answer these two first
These two answers are the "done?" check from the loop.
Write it so it could be checked without you. "Every overdue invoice got one reminder." Not "chase the money."
Different jobs need different checks. Pick the one that fits ↓
Anthropic released a free skill called launch-your-agent that builds it for you through an interview. You answer plain questions, it does the rest. The recipe is always the same three things. Here is the shape, filled in for chasing payments:
One clear sentence. "Remind overdue clients." Not "fix my cash flow."
→ that's the GOAL boxGive it a finish line it can verify, plus a hard stop so it can't run forever.
→ that's the DONE? checkSee where it trips. Fix the instructions, not the output. Then let it run on its own.
→ that's reason → act → observeMost agents fail for boring reasons: they run forever, burn money, or ship junk. The good ones get these right.
Define "done" so it could be verified, not "make it good."
every overdue invoice emailedA max number of tries or a budget. Always, so it can't run forever.
stop after 100 sitesConnect it only to what the job needs: your inbox, calendar, files.
email · spreadsheetIt remembers past runs, so it gets better and never repeats itself.
who was already remindedFor work that matters, a second agent grades it. It never grades itself.
does this match my voice?Big multi-step job? Have it outline first. Small one? Skip it.
outline before actingAnything client-facing pauses for your yes before it goes out.
approve before sendingAgents run on your own AI account. Start small and bounded, then scale.
watch the first runsThe Anti-Overwhelm Journal: one small, practical AI step you can actually use, in plain English. No jargon, no hype. For owners who would rather run their business than babysit their software.