This series has spent ten posts on what a hundred-thread team can do. A fair reader is owed the other post: the one about where it cannot.
The whole model rests on one move. Work gets handed to agents through a contract: what done means, what evidence proves it, what the agent may invent, where it must stop. Every gain in this series, the parallel calls, the swarms, the overnight threads, is a place where that handover worked.
So the line is not technical, and it is not about model quality. The line is whether the contract can be written. Where you can say what done means before the work starts, fleets scale. Where you cannot, they break, and the way they break is quiet, expensive, and confident.
Where it breaks
1. The first version of anything
You cannot write a contract for an unknown. The first version of a product, a design, a company voice, a novel: it is judged by a feel that nobody can specify yet, because the specification is what making it produces.
Ask an agent for "a good logo" and you get forty competent logos and no closer to the one you wanted, because "the one you wanted" does not exist in words until you see it. This is why infinite generation makes weak taste harder to hide: the fleet can generate the options, but the selection is the work, and selection cannot be delegated to the thing that generated.
The fleet scales the second version, and the tenth. The first version is yours.
2. Work that is one long line of thought
Some work is not a bundle of tasks. It is one continuous judgment: a negotiation, a founding decision, an architectural choice, an essay with an argument. The thread model wants to split work into pieces. Split this kind of work and you do not get parallel pieces. You get wreckage.
A fifteen-year-old knows this without a framework: you cannot write a group essay by having five people each write a sentence. The essay is not the sentences. It is the one mind that knows where the argument is going.
Agents can help along that line of thought, as a sounding board, a drafter of options. The moment the work itself needs one context held whole, it is a one-thread job, and pretending otherwise produces confident fragments.
3. Irreversible decisions
The trust ladder has a top, and some work never leaves level zero. Decisions that cannot be undone, firing, acquisitions, public statements of principle, safety-critical sign-offs, do not scale by delegation, not because agents are bad at them but because accountability is the work. The person who answers for the decision cannot be a system. A fleet can prepare the brief. It cannot stand in the room.
4. Trust and relationship
A thousand calls a minute is reach. Concurrency is not relationship. Every context where the asset is trust between specific humans, key accounts, partnerships, a community, a team's own morale, resists the fleet, not at the edge cases but at the center. The village does not parallelize.
5. Work with no ground truth
Evals are the engine of the whole model: exams for software, run on every change. But an exam needs right answers. In brand-new territory, where nobody knows what good looks like yet and the feedback arrives in months rather than minutes, there is nothing to grade against. Agents will happily produce confident output, and there is no exam to catch it. This is the swamp end of the bottleneck map: when verification cannot exist, generation speed is just the speed at which you get lost.
6. Work cheaper than its contract
The model has an overhead rate. Contracts, evals, logs, and reviews cost real human time, and some work does not clear the bar. Do not build a fleet to write a birthday card.
The honest math: if the contract takes an hour to write and the task takes ten minutes, the task is not automatable at this maturity, whatever the benchmarks say. This is where small, trivial, occasional work lives, and there is nothing wrong with a human doing it. Not everything worth doing is worth a system.
7. Work that changes shape faster than contracts age
Contracts encode this month's understanding of the task. Some domains rewrite themselves monthly: model capabilities move, markets move, the regulations move. If the task's shape changes faster than contracts can be updated, the fleet runs on stale instructions, and stale instructions with machine consistency are worse than human sloppiness, because the human would have noticed the world changed. This is weather, and fleets do not sail weather well.
8. Teams without the steering layer
The last break is not in the work. It is in the org. Uncontracted parallelism is chaos with extra steps, and a team that has no contract-writing, evidence-reading, escalation-owning discipline will multiply its mess. The METR trial found experienced developers 19 percent slower with AI tools, early 2025, and the lesson generalizes: capability added beyond an organization's steering capacity is a tax, not leverage.
When it scales
Turn the breaks around and the scaling conditions write themselves. The fleet earns its keep where most of these are true:
- [ ] The task repeats. Volume pays for the contract, and the tenth run costs almost nothing.
- [ ] Done can be named and checked. Someone can write the exam before the work starts.
- [ ] Mistakes are reversible. A wrong output costs a correction, not a funeral.
- [ ] The pieces are independent. No single mind needs to hold the whole for the pieces to make sense.
- [ ] Ground truth arrives quickly. Reality grades the work fast enough to steer by.
- [ ] The shape is stable. The task changes slower than contracts can be updated.
- [ ] A named human owns the outcome and reads the evidence. The steering layer exists.
Notice that none of these conditions mention model quality. The line is drawn by the shape of the work and the maturity of the team, which is why the same fleet that transforms one department is a toy in the next one.
The audit, in five questions
Point any piece of work at these and listen for a shrug:
- Can I describe what done means, before starting, so specifically that a stranger could check it?
- If this output is wrong, what does it cost, and who notices first?
- Does finishing one piece tell me anything about the next piece?
- Has this task's definition changed in the last quarter?
- Who owns the result: one name, or a shrug?
Shrugs on one and two mean level zero, the work stays human, and that is the correct answer, not a failure of imagination. All clear answers mean the work is fleet-shaped, and some systems assume exactly that surplus of scale and stability.
One more break, going up
The model also breaks at the top, and it breaks the same way every system does: coordination becomes the product. A fleet large enough needs a registry of owners, a schedule of eval maintenance, a review load, a budget process. Past some size, the fleet's own management is a bureaucracy that eats the leverage, and the thousandth thread costs more than it returns. Scale is not a victory lap. It is the next constraint, wearing a crown.
FAQ
Does not mean agents are overrated?
It means they are shaped. The fleet is a lever, and levers move specific things. The teams disappointed by agents were mostly sold a lever and handed a sphere: work with no contractable done, no volume, no ground truth. The tool was fine. The fit was never checked.
We tried agents on creative work and it failed. Does this explain that?
Mostly, yes. Drafts and options are fleet-shaped, and the fleet earns its place there. Selection, direction, and the taste that says this one, are the first-version judgment that cannot be contracted. The teams that fail try to delegate the selection too. The teams that work let the fleet fill the room and keep the choosing human.
Where is the line for a small early-stage team?
Blurred, and that is the honest answer. An early-stage team lives mostly on the break side, first versions, shifting shapes, no ground truth, which is why founder hours still matter most. The fleet enters through the edges: tests, support macros, research scans, the repeating perimeter. As the product stabilizes, the contractable center grows, and the fleet grows with it.
What is the one-sentence test?
If you can write the contract, the fleet scales it. If you cannot, the work is yours, and that is not the machine's failure. It is the reason your role exists.
Related reading
- Ten People, One Hundred Agents
- From Pull Requests to Proof
- The Bottleneck Keeps Moving
- More Agents Do Not Create Continuous Improvement
- The Swarm Is a Different Animal
- Judgment Does Not Transfer
- Infinite Generation Makes Weak Taste Harder to Hide
If you want the line drawn through your own work before the fleet is built, book a discovery call.