Most agent diagrams treat planning, verification, and repair as free intelligence. They are not. Every extra role consumes the same budget the system needs to solve the problem.

Researchers Thomas Nolasque, John Grey, Calista Pham, and Ankit Vani tested that tradeoff across 28,000 completed runs on two financial-reasoning benchmarks. They compared one GPT-5.4 mini call with a verified-search architecture that added planning, label-blind checking, and repair. Output-equivalent budgets ranged from 250 to 42,000 tokens.

At the smallest budgets, neither system could fit a complete prompt. At 1,000 tokens, the single call reached 18 percent accuracy while the structured system remained near zero because planning consumed the room needed for an answer. Between 1,000 and 1,500 tokens, the result flipped. Above that threshold, verified search stayed ahead, reaching roughly 44 percent at the largest tiers versus about 40 percent for the single call.

The exact crossover is not a universal law. It came from two financial datasets, one model, and one agent design in a new preprint. Change the task, context size, model, prompt, or checker and the economics move. What matters is the shape of the result: coordination overhead is fixed before it becomes useful.

That has direct product consequences. A workflow that wraps every easy request in a committee can become slower, more expensive, and less accurate. A workflow that never checks hard work saves tokens until one bad answer reaches a customer. The correct architecture is conditional. Route simple tasks directly, reserve search and repair for work where the error cost justifies the overhead, and measure the boundary.

Agentic software is not automatically better because more boxes appear in the diagram. The system earns complexity only when the additional passes change outcomes enough to pay for themselves. Intelligence has a budget. Architecture decides whether that budget becomes leverage or meeting overhead for machines.

LaunchPad positionAgent architecture has a fixed coordination cost. Teams should budget planning and verification as part of the task, then measure where that overhead begins to produce better answers instead of assuming more steps always mean more intelligence.
Reporting standard

This report draws on the linked primary sources and reputable reporting. Company statements are treated as claims until independently demonstrated.