The production agent stack: responses, tools and traces

Reliable agents need more than a model call. They need controlled tools, observable execution and evaluations tied to the workflow’s real outcome.

Responses, tools and traces work as one system

The agent stack is consolidating around three ideas: a response loop that can reason across steps, tools that connect the model to real systems and traces that show how the result was produced. OpenAI’s Responses API and Agents SDK are one expression of that pattern.

The model is only one subsystem

The model is only one part of the system. Retrieval determines which evidence enters the context. Tool definitions determine what the agent is allowed to do. State determines what survives between steps. Tracing and evaluation determine whether the workflow can be trusted after it leaves a prototype.

Test the workflow, not only the prose

This changes how teams should test AI. A polished answer is not enough. Tests should cover tool selection, permission enforcement, recovery from partial failure, source quality, latency and the point at which a human must take over. The evaluation set should resemble the messy edge cases of the real process rather than a clean demonstration script.

Start narrow, but make the first workflow complete

The best first deployment is therefore narrow but complete. Choose one workflow, connect the minimum tools it needs, trace every decision and compare the result with a known baseline. Expansion should follow measured reliability, not model enthusiasm.

Observability is part of the product

If an operations team cannot explain an agent’s path, it cannot responsibly improve it.