loading…
Architecture is choosing the least complex system that can satisfy the requirement.
The central question is not "Which agent framework should we use?" It is "Who should control the next step: application code, a model, a human, or another agent?"
Every layer of autonomy adds flexibility, but also expands latency, cost, security exposure, and the number of behaviors you must evaluate. Start with a direct model call and add retrieval, workflow logic, tools, or agents only when a concrete requirement demands them.
One model call → add retrieval when private/current knowledge is required → add workflow when the steps are known → add an agent when the path must be chosen at runtime → add multiple agents only when real specialization or boundaries justify it
| Axis | Question | Examples |
|---|---|---|
| Knowledge access | Where does evidence come from? | Prompt only, long context, RAG |
| Control flow | Who chooses the sequence? | Application workflow, model-driven agent loop |
| Actor topology | How many decision-making actors exist? | Single model, single agent, multi-agent |
RAG, workflow, and agent are not mutually exclusive labels. You can build a deterministic RAG workflow, a RAG-enabled agent, or a multi-agent system that shares a retrieval service.
Before adding a component, name the failure it solves and the new failure modes it creates. A planner may solve unknown task order but introduce loops. A second agent may improve domain isolation but create handoff errors. Architecture is disciplined trade-off management.