Openai launches new tools to help companies build artificial intelligence agents

On Tuesday, Openai has released new tools designed to help developers and institutions build artificial intelligence agents – automatic systems that can independently achieve tasks – using the company’s special artificial models and frameworks.

The tools are part of the new API from Openai, which allows companies to develop dedicated AI agents who can conduct searches on the web, scanning through the company’s files, and moving at web sites, just as it is very similar Openai operator product. API is effectively replaced by Openai’s responses API assistantsThe company plans to sunset in the first half of 2026.

The noise has grown around artificial intelligence agents in recent years despite the fact that the technology industry has struggled to show people, Or even specifyWhat are the “artificial intelligence agents” really. In the latest example of the pre -benefit working noise, the Chinese starting butterfly effect has become a viral earlier this week For a new AI agent platform called Manus Users did not quickly discover that many of the company’s promises.

In other words, the risks are high for Openai to obtain the agents properly.

“It is very easy to offer your agent,” Olivier Godmont, head of API products in Openai, told Techcrunch in an interview. “To expand the agent’s scope is very difficult, and to make people use it often very difficult.”

Earlier this year, OpenAi presented artificial intelligence agents in Chatgpt: The operator, which moves on web sites on your behalf, and Deep searchWhich collects research reports for you. Both the tool offered a glimpse of what the parent technology could achieve, but they left somewhat required in the “Self -Government” section.

Now with API responses, Openai wants to sell access to components that operate artificial intelligence agents, allowing developers to build deep search factors applications. Openai hopes that developers can create some applications with the agent technology that you feel more independent than available today.

Using the responsibility programming interface, developers can take advantage of the same artificial intelligence models (in the inspection) under Openai’s cover Chatgpt search Internet search tool: GPT-4O and GPT-4O Mini Search. Web browsing models can get answers to questions, citing sources because they generate responses.

Openai claims that the GPT-4O and GPT-4O MINI search are largely accurate. On the company’s simple standard, which measures the ability of the forms to answer short questions that seek the facts, the GPT-4O search results record 90 % while GPT-4O Mini Search Scarts 88 % (the highest better). For comparison, GPT-4.5 The larger model recently released by Openai – it records only 63 %.

The fact that the Acting Search tools are more accurate than traditional artificial intelligence models are not necessarily surprising-theoretical, the search for GPT-4O can search for the correct answer. However, searching on the web does not make hallucinations a problem of solutions. Besides its realistic accuracy, search tools of artificial intelligence also tend to struggle with short inquiries and navigation (such as “Lakers Score Toy”), and modern reports indicate this Chatgpt quotes are not always reliable.

API responses also include a file search tool that can quickly scan files via files in the company’s databases to recover information. (Openai claims that he will not train models on these files.) In addition, developers who use API responses can click on the Openai computer use model, which operates the operator. The model creates mouse and keyboard procedures, allowing developers to automate computer use such as data entry and application function.

Openai said that institutions can operate the CUA model, which is launched in the research inspection, locally on its own systems. The consumer version of the CUA available in the operator can not only take the web action.

To clarify this, API will not solve all the technical problems suffering from artificial intelligence agents today.

Although the AI’s research tools are more accurate than traditional artificial intelligence models-which is not surprising, given that they can search for the correct answer-do not provide the research on the web Amnesty International Helling is a solution problem. GPT-4O search remains 10 % of realistic questions. In addition to its accuracy, artificial intelligence research tools also tend to Struggle with short inquiries and navigation (Like “Lakers Score Toy”), modern reports indicate this Chatgpt quotes are not always reliable.

In a Blog for Techcrunch, Openai said that the CUA model “is not yet reliable to automate the tasks in operating systems”, and that it is subject to “unintended” mistakes.

However, Openai said that these are early repetitions of their agent tools, and they are constantly improving them.

Besides the API responses, Openai exports an open source tool called SDK agents, which provides developers free tools to integrate models with their internal systems, put guarantees, and monitor AI’s agent activities for error correction and improvement. SDK agents is a kind of some kind to Openai’s Swarm, a multi -agent coordination framework issued by the company late last year.

Judmont said he hoped that Openai will be able to bridge the gap between AI Agent offers and products this year, and this, in his opinion, “The agents are the most applications of Amnesty International that will happen.” This is the CEO of Openai Sam Altman in January: 2025 is the year in which artificial intelligence agents enter the workforce.

Whether in 2025 it became really “the year of artificial intelligence agent”, the latest Openai’s versions show that the company wants to switch from delightful fake offers to influencing tools.