a practical guide to building agents-OpenAI发布的构建智能体实用指南.pdf

Apractical


guideto


buildingagents

Contents

Whatisanagent?4

Whenshouldyoubuildanagent?5

Agentdesignfoundations7

Guardrails24

Conclusion32

2Practicalguidetobuildingagents

Introduction

Largelanguagemodelsarebecomingincreasinglycapableofhandlingcomplex,multi-steptasks.

Advancesinreasoning,multimodality,andtoolusehaveunlockedanewcategoryofLLM-powered

systemsknownasagents.

Thisguideisdesignedforproductandengineeringteamsexploringhowtobuildtheirfirstagents,

distillinginsightsfromnumerouscustomerdeploymentsintopracticalandactionablebest

practices.Itincludesframeworksforidentifyingpromisingusecases,clearpatternsfordesigning

agentlogicandorchestration,andbestpracticestoensureyouragentsrunsafely,predictably,


andeffectively.

Afterreadingthisguide,you’llhavethefoundationalknowledgeyouneedtoconfidentlystart

buildingyourfirstagent.

3Apracticalguidetobuildingagents

Whatisan

agent?

Whileconventionalsoftwareenablesuserstostreamlineandautomateworkflows,agentsareable

toperformthesameworkflowsontheusers’behalfwithahighdegreeofindependence.

Agentsaresystemsthatindependentlyaccomplishtasksonyourbehalf.

Aworkflowisasequenceofstepsthatmustbeexecutedtomeettheuser’sgoal,whetherthats

resolvingacustomerserviceissue,bookingarestaurantreservation,committingacodechange,


orgeneratingareport.

ApplicationsthatintegrateLLMsbutdon’tusethemtocontrolworkflowexecution—thinksimple

chatbots,single-turnLLMs,orsentimen