What are OpenAI's Responses API and Agents SDK, and how do they help build AI agents?

On March 11, 2025, OpenAI released the Responses API, a new way for apps to call its models that comes with built-in web search, file search and computer use tools, and the Agents SDK, a free open-source toolkit for coordinating one or several AI agents. Together they give developers ready-made parts for building agents, programs that carry out tasks on a user's behalf, instead of assembling those parts themselves.

What are OpenAI's Responses API and Agents SDK, and how do they…

Yesterday, Tuesday, March 11, 2025, OpenAI released a set of tools for developers who want to build AI agents. OpenAI defines an agent as a system that independently accomplishes tasks on behalf of users: not just answering a question, but looking things up, reading documents, clicking through a website and handing back a result.

The company already offers agents of its own in ChatGPT, Operator, which navigates websites for you, and deep research, which writes research reports. With yesterday's release, it is selling the building blocks behind such agents so other companies can make their own.

What OpenAI released

There are four pieces.

The Responses API. An API is the doorway programs use to talk to a service. Until now, most developers used OpenAI's Chat Completions API, which is simple: send messages, get a reply. The older Assistants API added tools and memory but was more complex. The Responses API combines the two: one request can use several tools and several rounds of model work before returning an answer.
Built-in tools. Web search, file search and computer use come ready to switch on.
The Agents SDK. An SDK, or software development kit, is a package of code that saves programmers from writing common parts themselves. This one is open source, free to use and modify, and it coordinates several agents working together.
Observability tools. Tracing lets developers replay each step an agent took, which is essential when something goes wrong.

The Responses API is open to all developers and has no separate fee; you pay the normal rates for tokens and tools.

The three built-in tools

Tool	What it does	Price at launch
Web search	Finds current information online and cites its sources, using the same models as ChatGPT search	From $25 to $30 per 1,000 searches
File search	Searches your own documents, such as a help center or a policy manual	$2.50 per 1,000 searches, plus storage at $0.10 per GB per day after the first free GB
Computer use	Lets the model see a screen and send mouse clicks and keystrokes	$3 per million input tokens and $12 per million output tokens; research preview for some developers

OpenAI reports that its search models answer 90% (GPT-4o search) and 88% (GPT-4o mini search) of the questions in SimpleQA, a set of short factual questions, correctly. TechCrunch points out the other side: about one answer in ten is still wrong.

The computer use tool runs on the Computer-Using Agent model that powers Operator. Unlike the consumer Operator, which only works in a web browser, businesses can point it at their own systems. OpenAI is frank about its limits: on OSWorld, a test of everyday tasks on a full computer, it succeeds 38.1% of the time, so OpenAI recommends that a person supervise it outside the browser. It has added checks against prompt injection, hidden instructions planted in a web page or document to hijack an agent, and asks for confirmation before sensitive steps.

How the Agents SDK organizes work

The Agents SDK replaces Swarm, an experimental framework OpenAI released last year. It rests on four ideas:

Agents: a model plus clear instructions and the tools it may use.
Handoffs: one agent passing the conversation to another better suited to it.
Guardrails: safety checks on what goes in and what comes out.
Tracing: a record of every step for debugging.

Picture a shop's customer service. A receptionist agent reads each message and routes it. A shopping agent with web search answers product questions. A support agent with permission to file refund requests handles returns. A guardrail blocks requests that have nothing to do with the shop. Here are the instructions a builder might give the receptionist:

Prompt

You are the first point of contact for a small outdoor-gear shop.
Read the customer's message and decide who should answer it:
- product questions go to the shopping agent
- orders, returns and refunds go to the support agent
If the message is unrelated to the shop, reply politely that you can only help with shop questions.

The SDK works in Python today, with Node.js support promised soon. It can sit on top of either the new API or Chat Completions, and even on other companies' models if they offer a similar interface.

Why it matters

Agents have generated more excitement than working products. Olivier Godement, who leads OpenAI's API product, put the gap simply to TechCrunch: demonstrating an agent is easy, but scaling one and getting people to use it often is hard. By offering search, document reading, computer control and coordination as standard parts, OpenAI is betting that more teams can move from demos to tools people rely on.

There is also a signal about the future of OpenAI's platform. OpenAI says the Responses API is the direction for building agents. Chat Completions stays fully supported. The Assistants API will be phased out once the new API matches its features, with a target shutdown in mid-2026.

Try it: Write down one task you would hand to a helper, such as "find three local caterers that serve vegetarian food and summarize their prices." Next to it, list which of the three tools the agent would need (web search, file search, computer use) and the one step where you would want it to ask you before going further.

Should you switch from Chat Completions to the Responses API?

If your app simply sends a prompt and shows the reply, there is no rush. OpenAI says Chat Completions remains its most widely used API and will keep receiving new models whenever they do not depend on built-in tools. Your existing code keeps working.

If you are starting something new, OpenAI recommends the Responses API, which it describes as covering everything Chat Completions does and more. It makes the most sense when your app needs current web information, answers drawn from your own documents, or several steps in a row. Adding web search, for example, takes a single line that names the tool.

If you built on the Assistants API, start planning. Nothing changes immediately, and OpenAI promises a migration guide that keeps your data. But the new API is where its agent features will grow, so new projects are best begun there.