Not every AI capability needs an agent. A schema-bound intent is often the clearest way to expose extraction, classification, summarization, or controlled generation. Agent behavior becomes valuable when the capability must choose and use approved tools across multiple steps.
Keep the application contract stable
The consuming product should still call the intent it owns. Whether the runtime completes that work through one model call, retrieval, an agent loop, or a workflow is an implementation decision behind the contract.
Make orchestration observable
Tool choices, intermediate steps, failures, retries, and final output should be available as execution evidence. Without that record, complex behavior becomes harder to test and operate than the manual process it replaced.