AI Agents and Productivity Theater: 4 Principles to Make Them Do Real Work

Key takeaways

• More agents and more output is not the same as more done
• The real cost of agent work is often the human reading, checking and deciding that comes after
• Agents make experiments cheap, so explore widely but commit rarely
• Every recurring agent needs an end date or kill rule before it starts
• A manager agent should protect your attention, not hand you a longer to-do list

Agent swarms, parallel agents, AI “factories”: there seems to be a new phrase every few weeks for running more AI and producing more stuff. It is easy to feel productive watching it all happen.

In a recent video, AI strategist Nate B Jones calls a lot of this productivity theater. It looks like work and it burns tokens like work, but it often ends up somewhere nobody reads. His four principles come from how he manages his own agents in tools such as Claude and Codex.

Here is what he says, what we think it means for small businesses and solo creators, and two practical guides to help you put it into action.

What is AI productivity theater?

AI productivity theater is activity that looks like progress but does not lead to a result you care about. Think of the friend with a dozen AI projects, new agents running all day and a fresh plan every week, whose business still has not launched.

Nate’s point is that this is getting harder to spot, not easier. Agents can generate reports, prototypes and plans in minutes, so the volume of plausible work keeps rising. That makes being intentional about what you ask for more important than it has ever been.

What are the four principles for managing AI agents?

Nate’s four principles are:

1. Count the human work that comes after the agent finishes
2. Run more experiments than you commit to
3. Decide when an assignment ends before you make it recurring
4. Give the agent managing your work a budget for your attention

Why should you count the human work after the agent finishes?

Because the agent’s few minutes are only the first cost. Someone still has to read the result, understand it, decide what they believe and work out whether anyone cares. Share it with your team and they pay that cost too.

We tend to judge an AI result by how long it would have taken without AI: “this report would have taken a week”. Nate argues that tells you about the tool’s capability, not whether the report deserves your next ten minutes. You may already know enough to decide, or the decision may belong to someone else.

Turn that report into a weekly one and you have created a weekly claim on human attention, possibly for several people, from a job the agent finished in moments.

His positive example is a private chef in South Jersey, in the US, who used Claude to produce professional proposals, which he credits with bringing in around $20,000 in six weeks. Winning a proposal created more work, but it was the work the chef actually wanted: cooking for paying customers.

How do you make an agent’s output easier to check?

Ask the agent to show its working, and point your checking at the part that matters most. Nate describes a painting contractor near Chicago who uses Claude to read PDF blueprints floor by floor when preparing bids. On an early bid he caught it using a floor-area multiplier rather than the real wall measurements.

Now he asks it to show its calculations. His experience tells him which part of an estimate needs attention, so he shapes the assignment around that check. A more polished proposal would not have fixed the problem; a wrongly priced one would have been worse than useless.

You can also give the checking job to another agent: ask it to test the evidence behind a recommendation, find the assumption that would change the answer, or investigate something you would otherwise have to look into yourself. That is extra capability spent on making the human part easier, not on producing more.

Why run more experiments than you commit to?

Because agents make exploring cheap, but committing is still expensive. You can test an idea, try an approach and let it go, and learning that an exciting idea does not survive reality is valuable even if you keep nothing.

The trouble starts when every experiment feels obliged to become a project. Each project brings decisions, feedback and maintenance into your week. Starting is now so cheap that you can take on more responsibilities in a month than you could ever handle.

Nate’s example: a business where customers try the service once and do not come back. You could send several agents to investigate the same question from different angles:

• One reads the cancellation reasons you already have
• One looks at where customers ask for help
• One checks whether people who return buy something different from people who leave

Those investigations might point to different explanations, which lead to very different decisions. Compare that with asking one agent to design a loyalty scheme, another to rebuild the pricing page and another to write a win-back email. All three might come back looking good, but now you are reviewing three solutions before you understand the problem, and the sunk cost makes them hard to drop.

When do parallel agents actually help you focus?

Parallel agents help when each one explores a genuinely different angle on the same central problem. Your attention stays on the problem while the agents explore it, and you learn from each branch without living inside all of them.

They do not help when five agents produce five versions of the same answer. That leaves you a heavy comparison job that teaches you very little.

Nate also stresses noticing when the missing information is outside what agents can find. If the question is whether customers will pay, eventually someone has to ask them to pay. If the question is whether a tool helps people, you have to talk to people. Another simulation or polished draft can feel safe, and an agent may recommend it, but it will not answer that question.

When should a recurring agent task stop?

Decide when a task should end before you make it recurring. A recurring agent can easily outlive the reason you created it. Nate admits doing this himself: he asked for a weekly update on a new market, chose a different market two months later, and the update kept arriving unread.

The agent was doing exactly what it was told. As he puts it, that is not an agent problem; it is a planning problem on our side. Each small recurring job looks clever when you set it up, but together they fill your day with questions you no longer need answered and quietly use up your subscription or token budget.

• Ask whether it really needs to recur, or whether it is a one-off question to answer and move on from
• Set a trial length, such as “run for three days, then stop”, and restart it on purpose if you still want it
• Add a stop rule, such as “if nobody responds to this for three days, stop sending it”
• Stop when it stops teaching you: if the next pass adds nothing new, end it

He adds one more warning: do not let agents set the direction of your curiosity. Give them room to find something surprising, but you decide whether it is worth pursuing, even if you do not understand all the technology.

What is an attention budget for a chief of staff agent?

An attention budget is a clear brief telling your managing agent what deserves your time and what does not. Nate does not mean a literal number of minutes. He means the agent coordinating your work, often called a chief of staff agent, should understand your existing commitments and what a new assignment would add to them.

Research that was cheap for an agent can still imply a big decision for you, such as entering a new market. A good chief of staff agent should recognise that, and weigh it against what else you are doing, whether that is hiring, a launch or learning a new skill.

He suggests telling it what you care about. For example, if you value customer stories, it could surface an unusual customer comment you never asked for, because it fits your priorities. In his view, the agent should be cleared to interrupt you for:

• Commitments that need following through so nothing is missed
• New evidence that could change a decision you have made
• Choices only you can make, explained with what each option would require

Its job is to prepare decisions, explain what they need, help carry out work you have approved, and then let your attention move elsewhere. Most people get this wrong, he says, because they let the chief of staff surface everything.

Which questions show if your agents are doing real work?

Nate closes with a simple audit. For each agent you are running, ask:

1. Does it help with something I actually care about, or is it just running?
2. Is it an experiment I am still learning from, or is it only running because I forgot to stop it?

If you cannot answer, you may be in productivity theater. He also suggests a “truth-teller” style instruction that asks agents to say plainly when something does not matter or is only an experiment, rather than labelling everything important. His own version of that prompt is in his paid guide, linked from the video description.

What should small businesses do next?

• Start with the outcome, not the agent: one result you want this month, such as more booked calls or a launched product
• Limit live agents to the few tasks that feed that result
• Keep a simple agent log: task, owner, who uses the output, end date
• Review weekly and switch off anything nobody read
• Set approval points for anything that spends money or contacts customers (see How to Set Human Approval Boundaries for AI Workers)

This links back to something we believe strongly at Prodify Digital: simple systems you own and understand usually beat sprawling ones. A few well-briefed agents feeding your email list, your content and your products will do more than a dozen running on autopilot.

Frequently asked questions

The bottom line

AI agents can do a remarkable amount of work, but only you can decide which work matters. Count the human time after each run, explore freely but commit rarely, give every recurring task an ending, and make your managing agent guard your attention. Your next step: audit the agents you have running today and switch off one you no longer need.

Related posts

• How to Spot Reward Hacking When an AI Agent Says the Job Is Done
• How to Run Long AI Agent Jobs Overnight Without Wasting Money
• How to Plan for AI Automation in a Small Team Without Panic

Garry Knight

I'm Garry Knight, the person behind Prodify Digital. I write about email list building, email marketing, SEO, AI search and the tools that connect them. My aim is to make online marketing easier to understand, so creators and small business owners can make informed decisions about building an audience and keeping people engaged. Here you'll find straightforward guides and product reviews that explain what something does, where it fits and which limitations matter. The focus is on clear explanations and useful next steps—not hype, shortcuts or promises of easy earnings.

See Full Bio