A Practical Guide to Building Agents
Source metadata
- Publisher: OpenAI
- Format: Business and engineering guide (PDF)
- Landing page: OpenAI
- Canonical document: PDF
- PDF metadata creation date: 2025-04-07
Original summary
OpenAI defines agents as systems in which a language model manages workflow execution, selects tools, recognizes completion, corrects actions, and can return control to a person. It recommends agents for workflows involving nuanced decisions, difficult rule sets, or substantial unstructured data, while retaining deterministic automation where it is sufficient.
The guide organizes an agent around three components: a model, tools, and instructions. It recommends establishing an evaluation baseline with a capable model and then testing smaller models to reduce cost and latency. Tools are grouped into data, action, and orchestration functions, while instructions should convert policies and operating procedures into clear actions and account for edge cases.
For orchestration, the guide favors maximizing a single agent before adding multi-agent complexity. When specialization is needed, it describes a manager pattern, where one agent invokes specialists as tools, and a decentralized pattern, where peer agents hand control to one another.
The safety section treats guardrails as layered defenses combined with conventional authentication, authorization, access controls, and software security. It recommends risk-rating tools, validating inputs and outputs, and escalating repeated failures or high-risk actions to a human.
Important claims
- Agents are most useful where deterministic automation struggles with ambiguity, brittle rules, or unstructured information.
- Model selection should be evaluation-driven: establish an accuracy baseline first, then optimize cost and latency where smaller models remain adequate.
- A single agent with clear tools is often easier to evaluate and maintain than a premature multi-agent architecture.
- Overlapping or poorly differentiated tools can impair tool selection even when the total tool count is modest.
- Guardrails should be layered and supplemented by standard security controls rather than treated as a complete security solution.
- Human intervention is especially appropriate after repeated failures and before sensitive, irreversible, or high-impact actions.
Limitations and provenance notes
This is a vendor-authored practical guide that emphasizes OpenAI’s Agents SDK and reports generalized customer experience without providing a systematic evidence base for every recommendation. Its design advice should be evaluated against application-specific measurements.
The landing page presented a browser security challenge during access. The source was reviewed through the canonical PDF linked from OpenAI’s content delivery domain. The PDF metadata date identifies document creation, not necessarily public release.
The copyright status does not clearly permit republication of the guide, so this record preserves only metadata, an original summary, and links.