Start with runtime ownership
Agents API is a managed harness. Agents SDK is an application-run orchestration library. Responses is a model-and-tools interface for application-controlled workflows. These are architectural choices, not three packages that every project must install together. Compare runtime ownership before selecting surrounding infrastructure.
What your application still needs
Provide server-side authentication, user-to-session authorization, credential storage, and result delivery. A production app commonly persists its own job and ownership mapping. Managed conversation state does not replace your customer database or prove that the current user may access a session ID.
When compute is optional
A task using remote MCP and application functions can use no environment. Add compute when the task needs files, shell execution, or environment-local capabilities. Self-hosted compute requires a provider lifecycle controller and executor connection; it does not move the managed harness into your infrastructure.
When retrieval and queues help
Use retrieval when the task needs a searchable knowledge collection; a vector database is not a prerequisite for every agent. Use a durable inbox or queue when processing must survive the originating HTTP request. Choose these for workload requirements, not because they appear in another project's architecture diagram.
Trace one concrete workload
For an order-support workflow, the browser authenticates to your server; the server selects a session and grants an order-lookup function; the function checks ownership against your business database; the agent receives only the authorized result; the server returns progress and output to the browser. Add retrieval for policy documents and compute for file generation only when those tasks are needed.
Read the official reference
Check the source for current API fields, account requirements, and service limits.
OpenAI: architecture