Choosing the Right Architecture Under Constraints
Choosing an AI-agent architecture means deciding how much autonomy the workflow can justify, not selecting the most elaborate pattern available. This chapter applies four constraints—control, problem complexity, resources, and specialized expertise—to the patterns already considered, then shows how hybrid designs can preserve a simple path for routine work while escalating genuine edge cases. The practical outcome is a conditional starting design with an evidence-based reason to become more complex.
The same task can deserve different architectures at different points in its lifecycle. A customer-support system may begin with one agent and a small tool set, add routing when distinct request types become measurable, and later coordinate specialists only where a single context fails. That is not architectural indecision. It is a recognition that the design must fit the work actually being done, the risk of being wrong, and the resources available to operate it.
Four constraints decide the starting architecture
Control determines how much uncertainty is acceptable
The first question is how precisely the organization must explain and constrain each decision. High-control work includes regulatory compliance, financial transactions, and safety-sensitive operations. In these settings, a single agent with explicit tools or a sequential workflow may be preferable because the chain of responsibility is easier to inspect. A decision that can be reconstructed by one owner and a predictable sequence is easier to challenge, correct, and govern than one assembled from several opaque interactions.
Moderate-control work such as customer support, content creation, or ordinary data analysis can tolerate more adaptive routing when a person can review exceptions and the consequences are recoverable. Lower-control work such as exploratory research or brainstorming can gain from collaboration because discovery and breadth matter more than a single deterministic path. This is not a claim that low-control work is risk-free. It means the organization is willing to accept more emergent behavior in exchange for exploring possibilities that are difficult to plan in advance.
Control also includes the user's experience of responsibility. A central manager may preserve one visible owner, while a handoff may deliberately let a specialist take over. A parallel workflow may produce several perspectives but require an explicit decision owner. The architecture should make that ownership visible. If no one can say who is accountable for the final action or the unresolved exception, the system has distributed authority without distributing responsibility.
The contrast is practical. In a compliance workflow, a reviewer may need to explain which rule, record, and approval stage produced a recommendation. Predictable sequence and limited autonomy may be more valuable than broad exploration. In a research workflow, the system may be allowed to pursue tangential leads and compare competing interpretations, provided the output is clearly presented as analysis rather than an irreversible decision. The same model capability can serve both settings, but the acceptable architecture changes because the cost of an unexplained transition changes.
Complexity and domain breadth determine whether one mind is enough
A single-domain task with an established procedure may need a single agent equipped with specialized Skills rather than multiple agents.