When to use completions

Choose a direct completion when one model call is the complete workflow and application code already owns the input, output, and control flow.

Good fits

• Summarize one document or record.
• Rewrite or translate content.
• Classify one input.
• Generate a short draft.
• Verify provider credentials and model behavior.
• Power a small internal utility without conversation state.

Choose the runtime shape

Need	Use
One free-form model response	Direct completion
One schema-validated model response	Structured output
Reusable instructions, tools, memory, or multiple turns	Agent
Repeated typed extraction from existing content	Extractor
Several deterministic and model-driven stages	Pipeline

Know the boundary

A direct completion makes one provider call. It does not:

• execute requested tool calls;
• load or save conversation memory;
• retrieve dynamic context automatically;
• repeat model turns;
• persist runtime events.

Keeping this boundary narrow makes completions easy to test and replace. Move to an agent only when the workflow actually needs orchestration.