Openai launches new tools to help business build AI agents

On Tuesday, Openai launched new tools designed to help developers and businesses build agents of AI – automated systems that can independently perform tasks – using the company’s own AI models and frames.

The tools are part of the new answers to Openai’s answers that allows business to develop personalized AI agents that can search the web, scan through company files, and navigate on websites, similar to similar OpenaiS API to the answers effectively replaces Openai API assistantswhich the company plans to sunset in the first half of 2026

Hype around AI agents has grown dramatically in recent years, despite the fact that the technology industry is struggling to show people or even defineWhat “AI agents” are actually. In the most example of an agent Hype, which moves in front of usefulness, the Chinese butterfly starting effect earlier this week became viral For a new platform for AI agent called Manus The fact that consumers quickly discovered does not fulfill much of the company’s promises.

In other words, bets are high for Openai to fix agents.

“It’s quite easy to demonstrate your agent,” Olivier Godemont, Openai’s API product, said in front of TechCrunch. “Scaling an agent is quite difficult to make people use it often is very difficult.”

Earlier this year OPENAI presented two AI agents in Chatgpt: An operator who navigates the websites on your behalf and Deep researchwho compiles research reports for you. Both tools offered an idea of ​​what agent technology could achieve, but left a lot to be desired in the Autonomy Division.

Now with the API to the answers, Openai wants to sell access to the components that power AI agents, which allows developers to build their own agent applications in the style of operator and deep research. Openai hopes that developers can create some applications with their agricultural technology that feel more autonomous than what is being offered today.

Using the API of the answers, developers can touch the same AI models (in visualization) under the lid of Openai’s Chatting Web Search Tool: GPT-4O Search and GPT-4O Mini Search. Models can view the answers network for questions, citing sources as they generate answers.

Openai claims that the GPT-4O Search and GPT-4O Mini Search are very actually accurate. In the reference indicator of the company, which measures the ability of models to answer short questions, search-looking, GPT-4o search search is 90%, while the GPT-4O Mini Search estimates 88% (higher). For comparison, GPT-4.5 – Much more, the recently released model of Openai – only estimates 63%.

The fact that the AI ​​search tools are more accurate than the traditional AI models is not necessarily surprisingly theory, GPT-4O search can simply seek the right answer. However, web search Does not make hallucinations solved problemS Beyond their actual accuracy, the AI ​​search tools also tend to fight With short, navigation requests (such as “Lakers Result today”), and the latter reports suggest that Chatgpt quotes are not always reliableS

API for answers also includes a file search utility that can quickly scan in files in the company’s databases to retrieve information. (Openai claims that he will not train models of these files.) In addition, developers using API of the answers can touch the model of Openai’s computer agent (CUA) that power the operator. The model generates mouse and keyboard actions that allows developers to automate tasks for using a computer such as entering data and work processes of applications.

Businesses can optionally manage the CUA model, which is released in a research review locally in their own systems, Openai said. The CUA user version available in the operator can only take action on the network.

To be clear, the API of the answers will not solve all the technical problems that have hit AI agents today.

While the search tools powered by AI are more accurate than the traditional AI-fact models that is not surprising given that they can simply look for the right answer-searching the network does not depict AI hallucinations solved problemS Searching for GPT-4o still receives 10% of factual questions wrong. Beyond their accuracy, AI search tools also tend to Fighting short, navigation requests (as “Lakers Score Today”) and the latter reports suggest that Chatgpt quotes are not always reliableS

In a blog post provided to TechCrunch, Openai said the CUA is “not yet very reliable for automating operating systems tasks” and is susceptible to making “inattentive” mistakes.

Openai, however, said these were the early iterations of their agent instruments, and this is constantly working to improve them.

Along with the API to the answers, Openai launches an open source toolkit called Agents SDK, which offers developers free tools for integrating models with their internal systems, putting precautions and monitoring AI activities to fix errors and optimization. SDK agents are tracking varieties to Openai’s Swarm, a multiangent orchestration frame that the company launched late last year.

Godemont said he hoped Openai would overcome the difference between AI Agent’s demonstrations and products this year and that “agents are the most influential application of AI, which will happen.” This sounds proclamation, CEO of Openai Sam Altman, did in January: This 2025 is the year in which AI agents enter the workforce.

Whether 2025 really becomes the “AI Agent Year”, Openai’s latest editions show that the company wants to move from Flashy Agent Demos to impactful instruments.