The Task-Shape Test: How to Tell If a Task Is Actually Agent-Shaped 

When an agent-driven social network opened its doors, 1.6 million agents registered, and most never completed a single task. Not because the tools were broken, but because nobody had a way to tell which work on their desk was actually shaped for an agent. That diagnostic gap, not model capability, is the real bottleneck.

Why does everyone own an agent and use none of them?

We've been noticing a pattern wherever agentic tools land: people get real intelligence handed to them and freeze in front of it. Not from lack of imagination, from lack of a way to size up the problem. Buying "a little thinking" or "a lot of thinking" per task is a genuinely new kind of decision, and nobody grew up with instincts for it. So the agent sits there, capable and idle, while the task gets done by hand, or not at all.

What actually makes a task agent-shaped?

Strip away the hype and there are four questions worth asking about any task before handing it to an agent, or a team of them.

Size. Does the task fit inside what one agent can hold at full quality, or does it spill past a single context window? A calendar slot fits easily. A few hundred documents might not.

Independence. Can the pieces be done without knowing what the others found? Reading unrelated documents splits cleanly across agents, since no reader needs to talk to another. Coding tasks split well or badly depending on how independently the codebase is organized.

Separation of concerns. Do parts of the task need a different "mind" than the rest, the reason a bank won't let the person who enters a payment also approve it? A critic who didn't write the draft. That's a signal for more than one agent, not because either lacks skill, but because objectivity needs distance.

Checkability. Is verifying an answer meaningfully cheaper than producing one? A test suite, an exit code, a source you can point to, something you can glance at and call right or wrong. This one quietly decides more than it looks like it should.

Why does checkability decide everything else?

A widely cited 2024 Stanford study makes the case. Researchers gave a cheap coding model one attempt per bug and it solved 15.9% of them, unremarkable next to the best single-attempt model of that era, at 43%. Given 250 attempts at the same bugs, that cheap model climbed to 56%, beating the frontier single-attempt score without changing the model itself. Pushed to 10,000 attempts, a correct answer existed somewhere in the pile for over 95% of problems.

The catch, for anyone weighing a multi-agent build: that coverage only turned into results where a mechanical checker could grade each attempt. Where there was no checker, and something had to pick the best answer out of the pile instead, results stalled around 100 attempts, no matter how many more were thrown at the problem. The right answer was in there. Nothing could tell which one it was.

Separately, Anthropic has reported that in a multi-agent research system it built internally, token spend explained roughly 80% of the difference between a good run and a bad one, more than prompt wording did. Its team of agents beat a single frontier model working alone by 90.2%. The lesson isn't "more agents, always." A team is how you spend more tokens usefully than one agent can hold, and that only pays off once the output can be checked cheaply.

When is the right answer no AI at all?

Some things on the desk are judgment calls, not information-processing problems: which candidate to hire, what to name a product, which direction the business takes. AI can research around a decision like that and hand you a wall to bounce ideas off. It can't supply the instinct of someone who has sat across from both candidates and felt the difference. That's the task to sit with yourself.

How do task shape and architecture line up?

FAQ

Is a multi-agent system always better than one agent? No. It's usually far more expensive, and the extra spend only pays off when something can cheaply verify each attempt.

Why did so many people abandon their agents after signing up? Most weren't given a task shaped for what an agent does well. The tool wasn't the obstacle; not knowing where to point it was.

None of this needs a bigger model. It needs four honest answers about the work in front of you, and the willingness to say "not this one" when human judgment is genuinely cheaper than a token bill.