OpenAI launches new tools to help businesses build AI agents

On Tuesday, OPENA developers and initiatives have published new tools designed to help create AI agents – automatic systems that can perform work freely – using the company’s own AI models and frameworks.

Equipment is part of the new response to Openai, which allows business to develop custom AI agents that can perform web searching, scan through company files and navigate websites, prefer a lot OperatorThe Reactions effectively replace Openai Helpful APIThe company planned sunset in the first half of 2026.

Despite the technology industry fought to show people, the hype surrounding AI agents has grown dramatically in recent years, Or even defined“AI Agents” is actually what. In the most recent examples of agent hype utilities, the Chinese startup butterfly influence has gone viral early this week For a new AI agent platform called Manus The users who have discovered quickly have not provided the company many promises.

In other words, the bet is more for the agents to get the agents properly.

“It is quite easy to demo your agent,” said Opini’s API product chief Olivia Godmont in an interview. “It is quite difficult to scale an agent, and people are often very hard to use” ”

Early this year, OpenAI introduces two AI agents Chatzipt: Operator, which navigates websites on your behalf and Deep researchWhich compiled research reports for you. Both equipment provided a glimpse of what agent technology could achieve, but left quite somewhat to be desired in the “autonomy” section.

Now with the reactions of API, OpenAI wants to sell access to components that give strength to agents, allowing developers to create their own operators- and deep research-style agent applications. Opeena hopes that developers can create some applications with its agent technology that feels more autonomous than what is available today.

Reactions using APIs, developers can tap the same AI models (preferently) under the hood of Openai Chatzipt search Web Search Tools: GPT -4O Search and GPT -4O Mini Search. Models can browse the web for answers to questions, quoting sources as they create sources.

OpenAI has claimed that GPT -4O search and GPT -4O mini search are extremely correct. In the company’s SimplKUA benchmark, which measures the ability to answer models of short, true -search questions, GPT -4O search score 90% score and GPT -4O mini search 88% (higher). In order to compare, GPT -4.5 – Opening is very large, recently published models – only 63%score.

AI-powered searching equipment is more accurate than the traditional tidal AI model that is not necessarily amazing-according to the theory, the GPT-4O search can only find the correct answer. However web search Hallucinations do not render a solution problemThe Beyond their true accuracy, AI searching tools Tend to fight Short, with the navigational query (eg “Lakers Score today”) and recent reports that Chatzept’s quotes are not always reliableThe

The response API also includes a file search utility that can quickly scan the files in a company’s database to restore data. (OpenAI claims that it will not train models in these files)) Using API reactions, developers can tap the OPNA’s computer-use agent (CUA) model, which gives strength to the operator. The model produces the mouse and keyboard verbs, allowing developers to automatically automatically automatically use data entry and app workflies.

Openi says enterprises can alternatively operate the CUA model, which is locally published by research on their own system, Openai. The CUA’s consumer version available in the operator can simply take action on the web.

Obviously, the reactions will not solve all the technical problems that are overwhelmed by AII agents today.

Although AI-driven searching equipment is more accurate than the Traditional Taid AI model-a fact that they can only find the correct answer — not render the Web Search AI Hallucination is a solution problemThe GPT -4O search still gets 10% truthful questions incorrect. Beyond their accuracy, AI searching tools also tend to Short, struggle with the navigational queries (Such as “Lakers scores today”), and recent reports mean it Chatzept’s quotes are not always reliableThe

In a blog post provided to TechCrunch, the OpenAI said that the CUA model is still not highly reliable to automate automated work on operating systems “and it is sensitive to” unknowingly “mistake.

However, Openi said that they were the initial repetition of their agent tools and it continued to improve them.

In addition to API reactions, OpenAI Agents are releasing an open-source toolkit called SDK, which developers provide free equipment to integrate models with their internal systems, protect and observe AI agent activities for debugging and optimization. Agents SDK is a follow-up of Openai’s jerk, it is a multi-agent orchestration structure that the company has released late last year.

Godmont said that he hopes that Open can meet the gap between AI agent Demos and products this year, and according to him, “agents are the most effective application of AI.” It echoes Sam Altman, CEO of a Declaration made in January: That 2025 years of AI agents entered the workforce.

2025 really becomes a “year of AI agent” or not, the latest release of the Openai shows that the company wants to transfer from the agent agent to dominant equipment.