GPT WORLD / INDEPENDENT CODEX FIELD MANUAL Source review · 2026-07-31

Retrospective edition / AGENTS

Name the owner of the agent loop

The controller must own turns, tools, state, approvals, stop conditions, and traces; specialists receive narrow tasks and typed results.

01 / DECIDE

Architecture decision

Use the Agents SDK when SDK-managed execution, handoffs, guardrails, sessions, or tracing simplify the application.

The controller must own turns, tools, state, approvals, stop conditions, and traces; specialists receive narrow tasks and typed results. Map input, output, state, and side effects as one observable path before deciding which layer owns name the owner of the agent loop.

02 / BUILD

Three-step implementation

  1. 01

    Frame the contract

    Write down the caller, data classification, success condition, timeout, cancellation, and ownership. Use the Agents SDK when SDK-managed execution, handoffs, guardrails, sessions, or tracing simplify the application.

  2. 02

    Build one narrow path

    Implement one end-to-end path with request correlation, typed state, and reversible failure handling. The controller must own turns, tools, state, approvals, stop conditions, and traces; specialists receive narrow tasks and typed results.

  3. 03

    Prove the outcome

    Turn acceptance into a repeatable fixture, contract test, or browser test. Traces and evals compare task success, tool errors, handoffs, latency, cost, and human intervention with the baseline.

03 / BOUND

Production boundary

More agents do not guarantee better results; add a specialist only after a single-agent baseline and clear ownership exist.

04 / PROVE

Acceptance evidence

Traces and evals compare task success, tool errors, handoffs, latency, cost, and human intervention with the baseline.

SOURCE / HTTP

Reproducible source probe

curl -fsSI 'https://developers.openai.com/api/docs/guides/agents' | sed -n '1,5p'

PRIMARY / OFFICIAL

Official OpenAI sources

Sources checked · 2026-08-10

  1. 01Agents SDK