\n<\/p>\n
On Tuesday, OpenAI released new tools designed to help developers and enterprises build AI agents \u2013 automated systems that can independently accomplish tasks \u2013 using the company\u2019s own AI models and frameworks.<\/p>\n
The tools are part of OpenAI\u2019s new Responses API, which lets businesses develop custom AI agents that can perform web searches, scan through company files, and navigate websites, much like OpenAI\u2019s Operator product<\/a>. The Responses API effectively replaces OpenAI\u2019s Assistants API<\/a>, which the company plans to sunset in the first half of 2026. <\/p>\n The hype around AI agents has grown dramatically in recent years despite the fact that the tech industry has struggled to show people, or even define<\/a>, what \u201cAI agents\u201d really are. In the most recent example of agent hype running ahead of utility, Chinese startup Butterfly Effect earlier this week went viral for a new AI agent platform called Manus<\/a> that users quickly discovered didn\u2019t deliver on many of the company\u2019s promises.<\/p>\n In other words, the stakes are high for OpenAI to get agents right.<\/p>\n \u201cIt\u2019s pretty easy to demo your agent,\u201d Olivier Godemont, OpenAI\u2019s API product head, told TechCrunch in an interview. \u201cTo scale an agent is pretty hard, and to get people to use it often is very hard.\u201d <\/p>\n Earlier this year, OpenAI introduced two AI agents in ChatGPT<\/a>: Operator, which navigates websites on your behalf, and deep research<\/a>, which compiles research reports for you. Both tools offered a glimpse at what agentic technology can achieve, but left quite a bit to be desired in the \u201cautonomy\u201d department. <\/p>\n Now with the Responses API, OpenAI wants to sell access to the components that power AI agents, allowing developers to build their own Operator- and deep research-style agentic applications. OpenAI hopes that developers can create some applications with its agent technology that feel more autonomous than what\u2019s available today.<\/p>\n Using the Responses API, developers can tap the same AI models (in preview) under the hood of OpenAI\u2019s ChatGPT Search<\/a> web search tool: GPT-4o search and GPT-4o mini search. The models can browse the web for answers to questions, citing sources as they generate replies.<\/p>\n OpenAI claims that GPT-4o search and GPT-4o mini search are highly factually accurate. On the company\u2019s SimpleQA benchmark, which measures the ability of models to answer short, fact-seeking questions, GPT-4o search scores 90% while GPT-4o mini search scores 88% (higher is better). For comparison, GPT-4.5<\/a> \u2013 OpenAI\u2019s much larger, recently released model \u2013 scores just 63%.<\/p>\n The fact that AI-powered search tools are more accurate than traditional AI models is not necessarily surprising \u2013 in theory, GPT-4o search can just look up the right answer. However, web search does not render hallucinations a solved problem. Beyond their factual accuracy, AI search tools also tend to struggle with short, navigational queries (such as \u201cLakers score today\u201d), and recent reports suggest that ChatGPT\u2019s citations aren\u2019t always reliable<\/a>.<\/p>\n The Responses API also includes a file search utility that can quickly scan across files in a company\u2019s databases to retrieve information. (OpenAI claims that it won\u2019t train models on these files.) In addition, developers using the Responses API can tap OpenAI\u2019s Computer-Using Agent (CUA) model, which powers Operator. The model generates mouse and keyboard actions, allowing developers to automate computer use tasks like data entry and app workflows.<\/p>\n Enterprises can optionally run the CUA model, which is releasing in research preview, locally on their own systems, OpenAI said. The consumer version of the CUA available in Operator can only take actions on the web.<\/p>\n To be clear, the Responses API won\u2019t solve all the technical problems plaguing AI agents today. <\/p>\n While AI-powered search tools are more accurate than traditional AI models \u2013 a fact that is unsurprising given they can just look up the right answer \u2013 web search does not render AI hallucinations a solved problem<\/a>. GPT-4o search still gets 10% of factual questions wrong. Beyond their accuracy, AI search tools also tend to struggle with short, navigational queries<\/a> (such as \u201cLakers score today\u201d), and recent reports suggest that ChatGPT\u2019s citations aren\u2019t always reliable<\/a>.<\/p>\n In a blog post provided to TechCrunch, OpenAI said that the CUA model is \u201cnot yet highly reliable for automating tasks on operating systems,\u201d and that it\u2019s susceptible to making \u201cinadvertent\u201d mistakes.<\/p>\n However, OpenAI said these are early iterations of their agent tools, and it\u2019s constantly working to improve them.<\/p>\n Alongside the Responses API, OpenAI is releasing an open-source toolkit called the Agents SDK, which offers developers free tools to integrate models with their internal systems, put in place safeguards, and monitor AI agent activities for debugging and optimization purposes. The Agents SDK is a follow-up of sorts to OpenAI\u2019s Swarm, a framework for multi-agent orchestration that the company released late last year.<\/p>\n Godemont said he hopes OpenAI can bridge the gap between AI agent demos and products this year, and that, in his opinion, \u201cagents are the most impactful application of AI that will happen.\u201d That echoes a proclamation OpenAI CEO Sam Altman made in January: that 2025 is the year AI agents enter the workforce.<\/a><\/p>\n Whether or not 2025 truly becomes the \u201cyear of the AI agent,\u201d OpenAI\u2019s latest releases show the company wants to shift from flashy agent demos to impactful tools.<\/p>\n<\/div>\n