On Tuesday, Openai released new tools designed to help developers and businesses create AI agents – automated systems that can complete independent duties – using the company’s models and frameworks.
The tools are part of OpenAI’s new API answers, which allows businesses to develop custom AI agents who can perform online searches, scan the company’s files and browse websites such as Openai operator. The API answers effectively replaces the API Assistants of Openai, which the company plans to sunbathing in the first half of 2026.
The advertising campaign around AI agents has increased dramatically in recent years despite the fact that the technology industry has struggled to show people or even determine who the “AI agents” are. In the latest example of the agent advertising campaign running in front of the utility, the Chinese BUTTERFLY EFFECT Earlier this week became viral for a new AI agent platform called Manus quickly discovered by users did not give many of the company’s promises.
In other words, bets are high for Openai to get agents correctly.
“It is very easy to demo your agent,” Olivier Godmont, head of API of Openai, said in an interview with TechCrunch. “The escalation of an agent is hard enough and the use of people to use it often is very difficult.”
Earlier this year, Openai introduced two AI agents to Chatgpt: Operator, who navigates the websites on your behalf and deep research, which draws up research reports for you. Both tools offered a look at what agentic technology can achieve, but left enough to be desirable to the “autonomy” section.
Now with the API answers, Openai wants to sell access to accessories that feed AI agents, allowing developers to build their own operator and deep research applications. Openai hopes that developers can create some applications with the technology of its agent that feel more autonomous than they are available today.
Using the API answers, developers can use the same AI models (in the preview) under the hood of the Openai: GPT-4O search and GPT-4O mini search. Models can browse the web for answers to questions, citing sources as they create answers.
Openai claims that the GPT-4O search and the mini search-4O are extremely accurate. In the company’s Simpleqa Benchmark, which measures the models’ ability to answer short Fact search questions, the GPT-4O 90% search ratings while GPT-4O mini search ratings (higher). For a comparison, GPT-4.5-much larger, recently released OpenAI model-is just 63%.
The fact that AI-powered search tools are more accurate than traditional AI models are not necessarily amazing-theoretically, GPT-4O search can simply look for the right answer. However, internet search does not make hallucinations a resolved problem. In addition to their actual accuracy, AI search tools also tend to struggle with short navigation questions (such as “Lakers Screw Today”) and recent reports indicate that Chatgpt reports are not always reliable.
The API of the answers also includes a file searching program that can quickly scan the files in a company’s databases to retrieve information. (Openai claims that it will not train models in these files.) In addition, developers using the API API can press the Agent-Agent-AGent (Cua) model of Openai, which dominates the operator. The model creates mouse and keyboard actions, allowing developers to automate computer use work such as data inserting and applications.
Businesses can optionally execute the CUA model, which releases in the preview of research, locally in their own systems, Openai said. The consumer version of the CUA available to the operator can only take action on the tissue.
To be clear, the API of the answers will not solve all the technical problems that are plaguing AI agents today.
While AI-powered search tools are more accurate than traditional AI-Shirt models that are no surprise, as they can simply look for the right answer-the internet search does not make AI illusions a resolved problem. The GPT-4O search still gets 10% of the actual questions. In addition to their accuracy, AI search tools also tend to struggle with short navigation questions (such as “Lakers Screw Today”) and recent reports suggest that Chatgpt reports are not always reliable.
In a blog post provided at TechCrunch, Openai said the CUA model “is not yet extremely reliable for automation of operational systems” and that it is prone to “involuntary” errors.
However, Openai said they were early repetitions of their agent’s tools and is constantly working to improve them.
Along with the API answers, Openai releases an open source tool called Agents SDK, which offers developers free tools for integrating models with their internal systems, creating safeguards and monitoring AI activities for the purpose of detecting errors. SDK agents are a monitoring of the items in the flock of Openai, a multi -agent orchestration framework that the company released in late last year.
Godmont said he hoped that Openai can bridge the gap between Demos and AI products this year and that, in his opinion, “agents are the most aggressive application of the AI that will happen”. This reflects a CEO of the Sam Altman Declaration made in January: That 2025 are the agents of the year and enter the workforce.
Whether in 2025 really becomes the “year of AI Agent”, Openai’s latest releases show that the company wants to shift from the agents’ fancy demos to aggressive tools.
