Every week someone posts that an AI agent runs their whole company. It answers the email, writes the content, runs the ads, manages the team. One prompt, total autopilot.
Here’s my take: the AI agents that actually work are boring.
They do one job. They run on a schedule or when something specific happens. They work from a written playbook. And before anything goes live, a person says yes.
Plenty of people building with AI already work this way. They just don’t make the most exciting posts.
Autopilot Is a Demo, Not a System
The “one agent does everything” video is real in the sense that someone recorded it. What you don’t see is the next hundred runs.
The research firm Gartner predicted in June 2025 that over 40% of agentic AI projects will be canceled by the end of 2027, “due to escalating costs, unclear business value or inadequate risk controls.” The same release says many vendors are “agent washing,” putting a new label on chatbots and old automation, and estimates “only about 130 of the thousands of agentic AI vendors are real.”
The companies building the models say the same thing in their own words. Anthropic, the company behind Claude, published a guide to building effective agents in December 2024. Its advice: “we recommend finding the simplest solution possible, and only increasing complexity when needed.” It also warns that “the autonomous nature of agents means higher costs, and the potential for compounding errors.”
Even the software companies selling agents point you toward small ones. ClickUp’s own help center tells new users to “start with Super Agents that perform specific steps or tasks,” like a daily report of tasks due that day.
Nobody who builds these things for a living is telling you to hand over the keys.
What I Built in ClickUp
Earlier this year I built a training demo in ClickUp for a made-up software company, to teach a marketing team how agents work.
One person creates one task: announce a new feature, and send the full content package. A lead agent reads the request, checks that it has what it needs, splits it into four jobs and hands each one to a specialist:
- Research: the market, the competitors, what the audience cares about.
- The brief: a creative brief built from the company’s own template.
- The copy: the actual content, written from the brief.
- The brand check: everything reviewed against the brand guidelines.
Everything lands as a draft for a person to review. One click, five agents, and no single agent trying to do it all.
Then I built helpers for a marketing team I lead, and gave each one a character so people would actually want to hear from it:
- Doc Brown nudges people when their time entries are missing. Getting the timeline right is the whole point of Back to the Future, so it fit.
- Dwight checks new tasks for a missing due date, a missing owner or a big job with no subtasks, and leaves a comment saying what’s missing. Rule enforcement with zero apology.
- Yoda reads the workspace once a week and suggests the three improvements worth making. Not ten. Three.
Some of them still run. Some I’ve turned off. The ones that earn their keep all have the same shape: one job, and someone who reads what they send.
The Boring Ones I Run Now
Most of the agents I run every week now are scheduled tasks in Claude. None of them would make a good demo video:
- A Monday ad report. It pulls last week’s ad numbers for one business and sends a short summary with a one-page PDF. It reads the ad account. It never touches it.
- A lead tracking check during a launch. Every morning for the first two weeks, it checked that each new lead still carried its ad tracking. If tracking breaks and nobody notices until the monthly report, you’ve lost a month of data. After two weeks it turned off, by design.
- A promotion sign-up count. Each morning during a promotion, it counts new sign-ups against the daily goal and suggests where to push.
- A Sunday page update. For a weekly event, it prepares the landing page change (next week’s date up top, last week’s removed) and waits for my OK before anything goes live.
- A test with its own undo. When I start an ad test, the end date, the weekly check-ins and the undo all get scheduled the day it starts. Nobody has to remember to turn it off.
- A Monday tidy-up of my own workspace. It proposes what to clean up. It never decides.
Here’s the funny part. By Anthropic’s own definitions, most of these barely count as agents. The guide describes “workflows” as “systems where LLMs and tools are orchestrated through predefined code paths,” and agents as systems that “dynamically direct their own processes and tool usage.” Most of mine follow a fixed path, with AI doing the reading and writing in the middle.
That’s fine by me. Gartner put it plainly: “Many use cases positioned as agentic today don’t require agentic implementations.”
Run It Like a Kitchen
Walk into any busy restaurant kitchen. The grill cook doesn’t make dessert. Every station does one job, from a written recipe, and nothing leaves the kitchen until someone checks the plate at the pass.
Nobody calls that inefficient. It’s how a kitchen gets through a full dining room on a Saturday without falling apart.
The agent that does everything is a kitchen with one cook. It works great right up until the dining room fills up.
This is the same whether you’re a dentist, a contractor or a restaurant. A dental office already splits the work: one person confirms tomorrow’s appointments, another handles billing, and the dentist signs off on the treatment plan. A good agent setup copies the org chart you already trust. It doesn’t replace it with one robot.
The Rules I Build By
- One job. If you can’t describe it in one sentence, it’s two agents.
- A clear trigger. A schedule, or one specific event. Not “whenever it thinks it should.”
- Written context. The playbook, the template, the brand rules and the list of past mistakes, in files it reads every time. An agent is only as good as what it can read.
- It reads freely and changes nothing live without a yes. Reports, drafts and flags are fair game. Publishing, sending to customers, spending money and changing settings wait for a person.
- An end date or an off switch. Some jobs only matter for two weeks. Put the end on the calendar the day you start.
- Someone reads what it sends. An agent whose messages nobody reads isn’t saving time. It’s adding noise.
Where I Build Them (As of October 2026)
Tools change monthly, so treat this section as a snapshot.
- ClickUp Super Agents live inside your ClickUp workspace, next to your tasks and docs. They can be triggered by an @mention, a direct message, assigning them a task, a schedule (hourly to monthly) or an automation. Every run uses AI credits. In ClickUp’s words: “This includes runs where the Super Agent only returns an answer and takes no other action.” ClickUp estimates 100 to 300 credits a run, at $0.001 a credit. By my math, an agent that runs every hour at 200 credits a run uses about 144,000 credits a month. Boring and narrow is also cheaper.
- Claude scheduled tasks come in two kinds. Local tasks run on your computer with access to your files, but only while the app is open and the computer is awake. Cloud tasks keep running when your laptop is closed, and are still labeled a research preview.
My rule of thumb: build the agent where its context already lives. If the work and the history are in ClickUp, start there. If they’re in a folder of files, start in Claude.
Spec Your First One
Before you build anything, fill out one page. If you can’t fill it out, you’re not ready to build it.
Here’s one filled out for a contractor’s office. It’s an illustration, not a client story:
- Job: Every weekday, list yesterday’s estimates that haven’t had a follow-up and draft a follow-up email for each one.
- Trigger: Weekdays at 7 AM.
- Context: The estimates export, the office’s follow-up wording, and the list of customers who asked not to be contacted.
- Output: A list of drafts in the office manager’s inbox.
- Approver: The office manager, who reads each one and hits send.
- Never: Send anything itself, change a price, or offer a discount.
- Broken if: No list shows up by 7:30, or the list is empty the morning after a busy day.
- Ends: A review after 30 days: keep it, change it or turn it off.
None of that requires a developer. It requires knowing your business well enough to write down one job clearly. That’s the part no agent can do for you.
The Bottom Line
The agent that runs your whole company makes a great post. The agent that shows up every Monday at 9 with the same report, drafted and waiting for you, makes a better business.
One job. A clear trigger. Written context. A person who says yes.
Start with the most boring thing you do every week. That’s your first agent.
If you want help picking the first job worth handing to an agent, let's talk.