Four Models Do Not Make a Factory

The appeal of multi-model orchestration is obvious. One model plans, another builds, another reviews, and a fourth resolves disputes. Weaknesses should cancel out. Work should spread across specialized machines. The system should become more capable than any one participant.

Then the boundaries arrive. Every agent needs the task, the relevant context, the decisions already made, and a definition of what its output must enable next. Each handoff can omit a constraint, flatten a disagreement, or turn a tentative guess into accepted fact. Add another model and you do not merely add capability. You add another translation surface and another place for the route to split.

That does not make orchestration a bad idea. It makes orchestration an engineering discipline. If the models can debate indefinitely, repeat the same work, overwrite one another's conclusions, or pass giant transcripts as context, the system has automated a committee. Committees are expensive even when nobody is on payroll.

Calling one agent a planner and another a reviewer does not specialize them. A real role has bounded authority, admissible inputs, a required output shape, and an exit condition. The planner may decompose the destination but may not approve its own assumptions. The builder may change allowed surfaces but may not redefine acceptance. The verifier may reject the result but may not quietly repair it and grade the repair.

Those contracts make disagreement useful. A verifier finding a failed condition is evidence. Two general-purpose agents arguing from the same transcript is token consumption. The system must know which artifact outranks which opinion, who owns an unresolved consequence, and when to stop machine discussion and route a decision to a human.

Human judgment still chooses the destination and owns the standards. It should not spend the day carrying messages between models, reconstructing which answer was current, or deciding whether another round of conversation might help. If orchestration creates a new coordination job, the machinery is unfinished.

The common answer to a confused agent team is to share more context. That can make the failure worse. A complete transcript carries stale plans, rejected approaches, speculative claims, and duplicated evidence alongside the facts. Later agents cannot reliably distinguish a settled constraint from a confident sentence that survived because nobody removed it.

A factory routes artifacts instead. The destination travels as an explicit specification. Architecture rules arrive from their authoritative location. A plan records assumptions separately from decisions. Implementation produces a change and an evidence trail. Verification receives the acceptance conditions and the result, not an invitation to inherit the builder's optimism.

This is the mechanism that lets multiple models help without forcing every model to understand everything. Each receives the smallest complete package for its role. Provenance survives the handoff. Rejected material does not keep voting. The factory retains state; the models perform work against it.

Running four agents at once feels like scale. It is only scale when their work can proceed independently and reunite under a mechanical rule. Asking four models to solve the same problem may produce useful candidates, but somebody or something must compare them against the destination. Otherwise the system pays for variety and selects by confidence, familiarity, or whichever answer arrived last.

The right comparison is not one model versus four models. It is the cost of producing a result that survives. Sometimes one strong worker with good tools and hard gates is the shortest route. Sometimes several agents exploring separate hypotheses or checking distinct failure modes win. The factory should route based on the shape of the work, not on a fixed belief that more voices mean more intelligence.

Our position is that model choice will keep becoming less important than this routing machinery. Models will change. Prices will move. A governed system can swap workers, measure which route succeeds, and retain the correction after a failure. A handcrafted council of favorite models has to be rediscovered every time one seat changes.

Start with the destination, then name the distinct transformations required to reach it. Give each transformation one owner, one input contract, one output contract, and one proof condition. Make retries finite. Put budgets on time and consumption. Preserve authoritative artifacts instead of conversations. Test the assembled result somewhere none of the builders controlled.

Then watch where a person still intervenes. If they are choosing a consequence, keep the gate. If they are copying context, settling avoidable duplication, rerunning a known check, or explaining which agent has the latest answer, encode the rule and remove the job. Repeatable coordination is still repeatable work, even when the workers being coordinated are machines.

Four models can outperform one. They can also fail four different ways while producing a larger bill and a more persuasive story. The difference is not team spirit. It is the factory: the contracts, routes, gates, budgets, and retained evidence that turn independent capability into one dependable result.