Skip to content

Should I use a completion or an agent?

Use a direct completion when one model call is the complete workflow. Use an agent when Anvia needs to manage reusable instructions, tools, memory, retrieval, hooks, or multiple model turns.

NeedPrefer
Summarize, rewrite, classify, or draft onceCompletion
Return one schema-validated objectParsed completion
Let the model choose and call toolsAgent
Continue a stored conversationAgent with memory
Retrieve context before answeringAgent with knowledge

Why not use an agent for everything?

A direct completion has a smaller control surface: one request, one provider call, and one result. An agent adds a bounded runtime loop and additional failure, latency, and usage paths. Add that orchestration only when the workflow needs it.

Does a direct completion execute tool calls?

No. A completion can return a model-requested tool call, but it does not run the handler or continue the model loop. An agent owns validation, execution, tool-result messages, and subsequent turns.

Does an agent own product policy?

No. The application still owns authentication, authorization, tenant scope, services, persistence policy, and the response exposed to users. Instructions guide model behavior; they do not enforce permissions.

See When to use completions and Agent runtime lifecycle for the full boundaries.

Built for Anvia.