On Tuesday, Openai issued new tools created to help developers and enterprises build agents – automated systems that can independently perform tasks – using the company’s own models and frames.
Tools are part of the new OpenAi responses API, which allows businesses to develop personalized agents of the one who can conduct online research, scan through company files and navigate websites, many like Openai’s operator product. API responses effectively replace the API of Openai assistants, which the company plans to set in the first half of 2026.
Hype about his agents has grown dramatically in recent years despite the fact that the technology industry has fought to show people, or even determine, what are the “agents of it”. In the latest example of Agent Hype running ahead of services, the Chinese start Butterfly Effect earlier this week went viral for a new agents called Manus that users quickly discovered did not deliver many of the company’s promises.
In other words, the shares are high for Openai to get the right agents.
“It is very easy to demonstrate your agent,” Olivier Godemont told Openai’s API product, for Techcrunch in an interview. “Scaling an agent is quite difficult, and making people use it often is very difficult.”
Earlier this year, Openai introduced two agents of him to Chatgpt: Operator, who sails on websites on your behalf, and deep research, which compiles research reports for you. Both tools offered a brief presentation of what technology agent can achieve, but left enough to be desired in the “Autonomy” department.
Now with API responses, Openai wants to sell access to the ingredients that empower him, allowing developers to build their own operator applications and deep research style. Openai hopes that developers can create some applications with his technology of agents that feel more autonomous than those available today.
Using the Answer API, developers can knock the same models of it (in the survey) under the hood of the Openai Chatgt search tool: GPT-4o Search and Request Mini GPT-4o. Models can browse online for answering questions, citing resources while they generate answers.
Openai claims that GPT-4o search and mini GPT-4o search are very accurate in fact. In the standard of the Simpleqa company, which measures the ability of models to answer short questions, fact search, 90% GPT-4o research results while 88% GPT-4o Research (higher is better). By comparison, GPT-4.5-Model much larger, recently released of Openai-only 63%.
The fact that energy searching tools are more accurate than traditional models of it is not necessarily surprising-in theory, GPT-4o research can simply require the right answer. However, web search does not make hallucinations a solved problem. Beyond their factual accuracy, he also tend to fight short, navigating questions (such as “Lakers Sodue Today”), and recent reports suggest that chatgt quotes are not always reliable.
The Answer AP also includes a file search tool that can quickly scan files in a company’s databases to get the information. (Openai claims that he will not train models in these files.) In addition, developers using API Answers can touch the model of OpenAi (CUA) computer use agent that empowers the operator. The model generates mouse and keyboard actions, allowing developers to automate computer use tasks such as data insertion and application work flows.
Enterprises can optionally run the CUA model, which is leaving the research, in place in their systems, Openai said. The consumer version available in the operator can only take action online.
To be clear, the API replies will not solve all the technical problems that the agents of it plague today.
Whereas the search tools with it are more accurate than the traditional patterns of it-a fact that is surprising given that they can simply seek the right answer-the online search does not make the hallucinations of it a solved problem. GPT-4o search still receives 10% of wrong factual questions. Beyond their accuracy, he also tend to fight short, floating questions (such as “Lakers Sodue Today”), and the latest reports suggest that chatgt quotes are not always reliable.
In a blog post given to Techcrunch, Openai said the CUA model is not “yet very reliable for automation of tasks in operating systems”, and that it is sensitive to make “unintentional” mistakes.
However, Openai said these are early repetitions of their agents’ tools, and is constantly working to improve them.
In addition to the API responses, Openai is emitting an open source vehicle called SDK agents, which offers developers free tools to integrate models with their internal systems, set protective measures and monitor the activities of it for debug and optimism purposes. SDK agents are a continuation of the types of Swarm of Openai, a framework for the multi-agent orchestration the company issued at the end of last year.
Godemont said he hopes Openai will overcome the gap between Demo and his agent’s products this year, and that, in his opinion, “agents are the most influential application of what will happen.” Echoing a pro -Lalamate CEO Openai Sam Altman made in January: Since 2025 it is the year of agents to enter the workforce.
Whether 2025 is really done or not “the year of the AI agent”, Openai’s latest omissions indicate that the company wants to move from the demonstrations of the influential agent in influential tools.
