Grab the One-Minute Test That Tells You If Your Task Needs a Chat, One Agent, a Team, or Nothing at All

This year more than 1.6 million AI agents signed up for a social network built only for agents. The overwhelming majority sat there, idle. Then, in China, a second market appeared. People paying to have someone uninstall OpenClaw, the free agent they’d rushed to install weeks earlier. A free tool, and a paid market for getting rid of it.

Of the two, the first is the one that stays with me. A million and a half agents, switched on and left to sit. Calling that a failure is generous, because failure needs an attempt, and most of these were never asked to do a single thing. Someone set each one up and never worked out what to point it at. They weren’t broken. They were never dispatched.

On Wednesday I published the case that the agents themselves are fine. A company of about two dozen of them rebuilt my wife’s website in an afternoon for roughly eight dollars, caught a fabricator, a cheater, and its own manager along the way, and shipped work that a ten-year accessibility professional called correct. That piece was about the machine, how to structure untrustworthy agents so their failures get caught by arithmetic instead of by you. Today is the other half, the half I get asked about more and have seen written about less: when do you turn the machine on?

Why does the question feel so hard? For most of history, more thinking meant one of two things. You hired someone, or you waited. Thinking came attached to people, and people are expensive, slow to find, and yes, supposedly, we all need sleep. That constraint broke. Now thinking is metered. Priced by the token, purchasable tonight, in any quantity, for a problem you only discovered this afternoon.

Nobody grew up with instincts for that. Nobody ever had to ask which task in their week deserved fifty dollars of purchased thought, because the question didn’t exist long enough for anyone to develop taste in it. When someone stands in front of a freshly installed agent and asks “what do I even do with this?”, the honest answer is that the question is a budgeting one. Imagination was never the thing in short supply. Our species has had about eighteen months to practice.

So this is a budgeting guide. Four things you can estimate about any task in about a minute, resolving into one of four verdicts: a chat, a single agent, a team, or don’t bother. That last one saves you the most money.

Here’s what’s inside:

• A guide, plus a tool to help you orient. Describe a task and the tool tells you whether it wants a chat, one agent, a team, or none of the above, and what to do next. The guide is the walkthrough behind it.

• The four estimates. Size, independence, separation, checkability, and how they resolve into a verdict.

• Why spending buys results, and where it stops. A Stanford paper brute-forced a cheap model from 15.9% to 56%, Anthropic found token spend explained 80% of the difference between good runs and bad, and two limits (the verification wedge and the context ceiling) sort every multi-agent pitch you’ll ever hear.

• Three tasks from my actual week, graded. A calendar problem, a forty-tool audit that paid for itself, and a judgment call the test refused to take.

• The economics of don’t-bother. Why a task can be perfectly agent-shaped and still not be worth building, and the two dials that decide.

By the end, you’ll size up any task on your desk in about a minute and know exactly where it belongs, before you spend a dollar finding out the hard way.