On Tuesday, Openai released new tools designed to help developers and businesses develop AI agents – automatic systems that could have been able to unite tasks – using its own AI models and company frameworks.
The tools are part of the new API's new responses, which allows businesses to develop custom AI agents who can conduct web searches, I -scan through company files, and navigate websites, similar to Openai of Openai's Product. API responses effectively replaced Openai's Assistants APIthe company's plan to sink in the first half of 2026.
Hype around AI agents has grown significantly in recent years despite the fact that the tech industry has struggled to show people, or even specifyWhat really is “AI agents”. In the latest example of the agent hype running ahead of the utility, the start of the Chinese butterfly effect earlier this week went viral For a new AI agent platform called manus Those users quickly discovered were not delivered to many of the company's promises.
In other words, the stakes are high for Openai to get the agents right.
“It's a bit easy to make your agent,” Olivier Godemont, head of the Openai product, told Techcrunch in an interview. “To measure an agent is a bit difficult, and to use people who are often difficult.”
Earlier this year, Openai introduced two AI agents Chatgpt: Operator, who navigates websites on your behalf, and Deep researchthat accumulates research reports for you. Both tools offer a glimpse of what the agent's technology can achieve, but left a little to wish the department of “autonomy”.
Now in the API responses, Openai wants to sell access to the power elements of AI agents, allowing developers to develop their own applications of the research agent agent. Openai hopes that developers can create some agent technology applications that feel more autonomy than today's available.
Using API responses, developers can tap both AI models (in preview) under the hood of Openai's Chatgpt search Web Search Tool: Finding GPT-4O and finding GPT-4O. Models can browse the web for answers to questions, citing resources as they form responses.
Openai says the GPT-4O search and the finding of the GPT-4O Mini is quite accurate. On the company's simpleqa Benchmark, which measures the ability of models to answer short, looking for reality questions, GPT-4o search scores are 90% while GPT-4O Mini 88% search scores (higher are better). For comparison, GPT-4.5 – Openai is larger, recently released model – only 63%marks.
The fact that AI-powered search tools are more accurate than traditional AI models is not necessarily surprising-in theory, finding the GPT-4O can only look for the right answer. However, the web search does not provide hallucinations -a solution of a resolved problem. Beyond their accuracy, AI search tools tend to fight short, navigation queries (such as “Lakers score today”), and recent reports suggest that Cotation Cotation is not always reliable.
API responses also include a file search utility that can quickly scan the files in a company's databases to obtain the information. (Openai says it does not train the models in these files.) In addition, developers who use API responses can tap a computer-useing agent (CUA) model of Openai, which operates the operator. The model generates mouse and keyboard actions, allowing developers to automatically use computer use activities such as data entry and workflows.
Businesses can optionally run the Cua model, which releases the research preview, locally on their own systems, Openai said. The Cua Consumer version available to the operator can only take web actions.
To be clear, API responses will not solve all the technical problems that are robbing AI agents today.
While AI-powered search tools are more accurate than traditional Ai-a-a-a-no surprise models that they can only find the right answer-the web search is not rendering Ai hallucinations -gula a resolved problem. The GPT-4O Search still gets 10% of real questions wrong. Beyond their accuracy, AI search tools are also likely to Struggle in short, navigation queries (such as “Lakers Score today”), and recent reports suggest that Cotation Cotation is not always reliable.
In a blog post provided in Techcrunch, Openai said the CUA model was “not yet fully reliable for automating activities on operating systems,” and it is easy to make “accidental” mistakes.
However, Openai said these were early iterations of their agent tools, and continued to work to improve them.
Next to the API responses, Openai releases an open-source toolkit called SDK agents, which offer developers free tools to combine models with their internal systems, put in protection areas, and monitor AI agent activities for the purposes of debugging and optimization. SDK agents are a follow-up of types in Openai's Swarm, a plot for the multi-agent orchestra released by the company last year.
Godemont said he hoped Openai could bridge the gap between the AI agent demos this year, and that, in his opinion, “agents are the most affecting AI application to happen.” That reflects a proclamation Openai CEO Sam Altman made in January: 2025 is the year of AI agents entering the workforce.
Indeed, 2025 has actually been the “year of the AI agent,” Openai's latest releases show the desire to move from strong agent demos to affecting tools.
