Build the Smallest Reliable Agent

Once a workflow has earned the right to be investigated as an agent, the problem changes from “Could a model help?” to “Which responsibilities must the system perform, and how will each be checked?” OpenAI’s agent guide reduces the foundation to three components: a model, tools, and instructions. This chapter explains how those components divide responsibility, why model choice should begin with an evaluated baseline, and where interfaces between them create failure.

The word smallest matters. A system does not become reliable because it has more tools, a more elaborate prompt, or a more powerful model. It becomes more understandable when each capability has a clear owner, each external action has a defined boundary, and the team can tell whether a failure came from reasoning, context, instructions, or the action itself.

Three components with different responsibilities

The model supplies judgment, but not the whole workflow

The model is the language model that reasons about the current state of the work and helps decide what to do next. Different tasks place different demands on that reasoning. OpenAI notes that a simple retrieval or intent-classification step may be handled by a smaller, faster model, while a harder decision such as refund approval may benefit from a more capable one. The trade-offs are task complexity, accuracy, latency, and cost.

The guide recommends starting a prototype with the most capable model for every task and using evaluations to establish a performance baseline. Only after the baseline is understood should the team try substituting smaller models where they still meet the accuracy target. This is not a permanent argument for the largest model. It is a way to avoid prematurely limiting the system and then confusing a model limitation with a problem in the workflow or instructions.

A baseline should answer more than “Does the response sound good?” It should test the decisions the workflow actually needs: whether the model identifies the right next step, selects an appropriate tool, handles missing information, and recognizes when work is complete. The guide does not prescribe a dataset, threshold, or cost formula, so the target must be defined for the particular workflow. A model that is adequate for a low-consequence draft may be inadequate for a decision that triggers a refund or changes a record.

This creates a boundary around model responsibility. The model can interpret context and propose a transition, but it does not automatically own the truth of the data it receives or the side effect of the tool it calls. Treating model output as the entire system hides the other sources of failure.

Tools connect the agent to information and side effects

Tools give the agent reach beyond its internal response. OpenAI groups them into three types.