Use the simplest architecture that works
A single LLM call is usually enough when one input produces one bounded output and no external action is required. An agent becomes useful when the system must choose among tools, inspect results, maintain workflow state, retry, or ask for approval.
Choose a simple model call when
- the task has a known input and output
- one prompt can contain the necessary context
- validation can happen immediately
- the system does not need to act in external systems
- deterministic code can handle the workflow around the model
Examples include classification, structured extraction, rewriting, and drafting from supplied context.
Consider an agent when
- the next step depends on an intermediate result
- several tools may be relevant
- work must pause and resume
- the workflow needs retries or escalation
- a human must approve selected actions
Do not use “agent” as a quality label
An agent adds latency, cost, permissions, state, and new failure modes. It is not automatically smarter than a well-designed workflow. Start with deterministic orchestration and grant the model only the decisions that cannot be encoded reliably.
A useful decision test is: what decision must the model make between steps? When there is no clear answer, a simple call or ordinary workflow is probably the better design.
Fact-check sources
Sources and product documentation can change. Recheck time-sensitive pages on the publication date.