Source: A Practical Guide to Building Agents (OpenAI)

OpenAI's 2025 guide on when to build an agent, agent components, orchestration patterns, guardrails and human-intervention triggers.

Contents

Summary

Distils customer deployments into guidance for a first agent: which workflows justify an agent, how to pick models, define tools and instructions, choose single- vs multi-agent orchestration, and layer guardrails with human intervention.

Key claims

  • Agents suit workflows that resisted deterministic automation: complex decisions with exceptions (refund approval), unwieldy rule sets (vendor security review), heavy unstructured data (insurance claims). If a use case does not clearly fit, a deterministic solution may suffice.
  • Model selection: prototype with the most capable model to set an eval baseline, hit the accuracy target, then swap in smaller models to cut cost and latency.
  • Tools come in three types: data, action and orchestration (agents as tools). Legacy systems without APIs can be driven by computer-use models.
  • Customers succeed more with an incremental approach than with an immediately fully autonomous architecture; maximise a single agent before splitting into several.
  • Plan for human intervention, especially early in deployment, to find failures and build the eval cycle. Triggers: exceeding failure thresholds (retries, repeated misunderstanding) and high-risk actions (cancellations, large refunds, payments).
  • Deployment is not all-or-nothing: start small, validate with real users, grow capabilities over time.

Feeds

AI use-case discovery · POC to production

Suggest changes →