Skip to content
Search

OpenAI

lastmod 2026-04-16created 2020-07-3133 edits

OpenAI is an AI research and deployment company based in San Francisco, California. Our mission is to ensure that artificial general intelligence benefits all of humanity. The OpenAI Charter describes the principles that guide us as we execute on our mission.

Categories

OpenAI Deep Research

AI Agent Tools

  • OpenAI releases developer tools for building AI agents | GeekNews
  • Web Search - Searches the web for up-to-date information and returns citation URLs
  • File Search - Semantic/keyword search over a list of uploaded files
  • Computer Use - Controls the computer and performs tasks
  • Responses API - Advanced unified response interface. Accepts text/image input and uses web/file search and CUA features
  • Agents SDK - Orchestration framework for agent development

OpenAI's practical guide to building agents

  • OpenAI's practical guide to building agents (32p PDF) | GeekNews
  • A practical guide for product and engineering teams building LLM-based agents for the first time
  • Unlike conventional automation systems, an agent is a system that makes its own judgments and takes action even in complex, ambiguous situations
  • Describes various design patterns, from single-agent systems to multi-agent orchestration
  • Emphasizes guardrail setup and human intervention strategies for reliability and safety
  • Recommends an approach that starts small and expands incrementally based on user feedback

What is an agent

  • An agent is a system that performs tasks on behalf of the user
  • Models that respond only once, such as simple chatbots or sentiment classifiers, are not agents
  • An agent has two core capabilities
    • Workflow execution management and decision making: judging completion, correcting errors, and returning control to the user on failure
    • Tool use: interacting with external systems to gather data or take action

When to build an agent

  • Suited to areas where traditional automation fails or is hard to maintain
  • Good use cases
    • Complex decision making: cases requiring nuanced judgment, such as approving customer refunds
    • Rule-based systems that are hard to maintain: cases with complex rules, such as automating security reviews
    • Unstructured data processing: cases requiring document analysis or natural language understanding

Basic elements of agent design

  • Model: the LLM responsible for reasoning and decisions
  • Tools: API functions that interact with external systems
  • Instructions: clear directions defining the agent's behavior

Model selection guidelines

  • Build the prototype with the best-performing model → then try swapping in smaller models with cost and latency in mind
  • Set up evaluation criteria (evals), verify accuracy, then optimize

How to define tools

  • Design tools in a standardized way to improve reusability
  • Tool types
    • Data retrieval: DB search, reading documents, web search
    • Taking action: sending email, updating CRM
    • Orchestration: calling other agents to divide the work

Best practices for writing instructions

  • Use existing operating procedures or policy documents to write clear step-by-step instructions
  • Break tasks down and specify them as concrete action units
  • Always include how to handle errors and exceptional situations (edge cases)

Orchestration patterns

Single-agent systems

  • One agent handles multiple tasks through a run loop
  • Uses prompt templates to handle various situations
  • Staying with a single agent is recommended as long as complexity is manageable

Multi-agent systems

  • Manager Pattern:
    • A central manager agent coordinates the workflow by calling several specialized agents like tools
    • Keeps a consistent interface with the user
  • Decentralized Pattern:
    • Agents hand off control to one another
    • An agent with a specific specialty takes control when needed
    • Suited to simple task triage early on

Guardrails

Purpose:

  • Protect data privacy and prevent system abuse
  • Maintain brand consistency and block inappropriate responses

Main types:

  • Relevance filtering: blocks off-topic input
  • Safety filtering: detects jailbreak attempts
  • Personal information (PII) filter: blocks exposure of sensitive information
  • Moderation: blocks violence and hate speech
  • Tool risk management: extra verification when calling high-risk tools

Building strategy:

  • Put data privacy and content safety first
  • Keep adding guardrails for new risks found in real use
  • Balance security and user experience

Human intervention (Human-in-the-loop):

  • Hand control over to a human when the agent fails or handles high-risk tasks
  • Typical triggers
    • Exceeding failure thresholds
    • Requests for highly sensitive tasks (refunds, payments, etc.)

Conclusion

  • Agents are innovative systems that can automate workflows in complex, uncertain situations
  • Build a strong model-tools-instructions foundation and expand orchestration incrementally
  • Guardrails and human intervention must be built in to ensure safety and reliability
  • Starting small, validating quickly, and expanding features incrementally is the key to a successful deployment

Two different approaches to fast LLM inference by Anthropic and OpenAI

See Anthropic#Two different approaches to fast LLM inference by Anthropic and OpenAI for details.

See also

Favorite site

References

  1. A-practical-guide-to-building-agents.pdf ↩

author Zer0 sitemap