OpenAI ResponsesAPI: Manus’s follower or standard setter?

Preface

On March 12, 2025, OpenAI released the Responses API and the open source framework Agents SDK designed specifically for the development of AI agents, marking Open AI's comprehensive upgrade from "dialogue generation" to "task execution".

Regardless of whether OpenAI admits it or not, I feel that this is clearly a response to Manus' recent popularity, and it also indirectly proves the correctness of Manus's intelligent agent route.

Publish content

First, let me introduce the content of this release.

Responses API

The Responses API is a set of development interfaces and an upgraded version of the Chat Completions API. It combines the ease of use of the Chat Completions API with the tool calling capabilities of the Assistants API, supporting developers to complete complex task chains through a single API request.

Its core tools include:

• Web Search Tool : Based on a fine-tuned GPT-4o or GPT-4o-mini model, it supports real-time retrieval of Internet information and automatically generates citation sources. The tool has an accuracy rate of up to 90% in the SimpleQA benchmark test and can be used in conjunction with other tools, such as combining file search to generate comprehensive reports.
• File search tool : supports multiple file formats such as PDF and Markdown, adds metadata filtering and custom sorting functions, and is suitable for retrieval enhancement generation (RAG) scenarios. Developers can quickly locate document content through the vector database, and integration only requires a few lines of code.
• Computer Operation Agent (CUA) : Based on the Operator model, the Computer Operation Agent (CUA) can simulate mouse clicks, keyboard input and other operations to automate browser or traditional software tasks. The tool has a success rate of up to 87% in tests such as OSWorld and WebArena, and provides environment isolation and sensitive operation confirmation mechanisms to ensure security.

Agents SDK

The Agents SDK was also open-sourced at the press conference, aiming to simplify the development, coordination and optimization process of AI agents.

It mainly includes the following core functions:

• Multi-agent collaboration : Based on the Swarm framework upgrade, it supports the orchestration of multi-agent workflows.
• Cross-model compatibility : Open source design allows integration of third-party models (such as Anthropic, Llama, etc.), and developers can freely choose the underlying LLM.
• Small amount of code development : Pre-configured templates and SDK assistants are provided, and complex tasks (such as file search + network search linkage) can be started with only a few lines of code.
• Enterprise-level expansion : Supports docking with traditional systems (such as legacy software without APIs) and achieves automated control through computer operation tools

Demo Case

An Agent case was demonstrated at the press conference, and the results looked very good.

The user name of the test simulation is Kevin, and he lives in Tokyo.

1. First, use the "File Search Tool" to understand Kevin's clothing preference data and generate a dressing style.
2. By calling the "Web Search Tool", store data near Tokyo is obtained and Patagonia stores are recommended.
3. Use "computer operation tools" to automatically operate web pages to place orders, including clicking, filling out forms, etc.
4. Finally, it also shows that the current agent pushes the return request to another customer service agent, allowing it to help the user complete the return.

Ambition

This time, OpenAI’s release is more of an application on the technical level. It may not look as cool as Manus, but I think there is a lot behind it.

For example, the Chat Completion API mentioned above may not be well known to those with a non-technical background, but it has actually become a standard feature of the big model API. Almost all mainstream big model APIs must be compatible with this specification, otherwise it will be very troublesome to use.

The Responses API is an upgrade of the Chat Completion API. It indicates OpenAI's ambition - perhaps all the intelligent agents we develop in the future need to follow this set of interface standards?

Conceptual distinction

The content of this release may be confused with several previous concepts, so a comparison is made here to facilitate everyone's understanding.

Responses API and Agents SDK

• Responses API : interface layer, providing tool calling and task orchestration capabilities, such as web search and file retrieval;
• Agents SDK : Framework layer, open source Python/Node.js library, supports multi-agent collaboration and process visualization monitoring.

The relationship between the two is similar to that between "Lego parts" and "assembly instructions", the former provides functional atoms and the latter defines the collaboration rules.

Differences between Responses API and Operators

• Operators : A specific agent (such as a browser plugin) based on the CUA model, focusing on end-user task execution;
• Responses API : The underlying support layer provides tool calling and model scheduling capabilities for operators and other intelligent agents

GPTs and Assistants API

• GPTs : Quickly deploy scenario-based AI services in a zero-code manner, similar to buttons in China.
• Assistants API : Provides enterprise-level conversation management, multi-role collaboration, and tool calling capabilities through standard interfaces.

Summarize

Compared with Manus's focus on rapid application implementation and user experience optimization, OpenAI's release this time demonstrates its deep understanding and advance planning for the future development of AI.

This reminds us that while pursuing application innovation, we should also enhance the accumulation and layout of core technologies and regulatory standards.

After all, DeepSeek has made a good start, hasn't it?