OpenAI has introduced a suite of new tools aimed at helping developers and businesses create more autonomous AI agents—systems capable of independently performing tasks such as web searches, file scanning, and website navigation. These enhancements are part of OpenAI’s newly launched Responses API, which replaces the Assistants API and is set to be phased out by mid-2026.
)
AI agents have become a focal point in the tech industry, yet their real-world utility remains a work in progress. Just this week, the Chinese startup Butterfly Effect made headlines with its AI agent platform Manus, only for users to quickly discover that it failed to live up to its bold claims. With mounting expectations and skepticism, OpenAI is under pressure to ensure its AI agents deliver tangible value.
>>>BW-C5 for Vivo iQOO Pad 3 Pro
A New Approach to AI Agents
OpenAI’s new Responses API builds on existing AI-powered tools such as Operator, which automates web navigation, and Deep Research, which compiles research reports. The goal is to provide developers with access to the underlying AI models that power these features, allowing them to build customized agent applications that feel more independent and useful.
The Responses API taps into GPT-4o search and GPT-4o mini search, the same models that power ChatGPT Search, OpenAI’s web search tool. These models are designed to browse the web and generate answers with cited sources. According to OpenAI, GPT-4o search scores 90% accuracy on factual questions, while GPT-4o mini search scores 88%—significantly outperforming OpenAI’s flagship GPT-4.5 model, which scores only 63%.
Beyond search, the API includes a file search utility, allowing AI to quickly retrieve information from company databases. OpenAI also introduced the Computer-Using Agent (CUA) model, the same system that powers Operator, enabling automation of computer tasks like data entry and app workflows. Enterprises will have the option to run the CUA model locally on their own systems for added security.
Challenges and Limitations
Despite these advancements, OpenAI acknowledges that its AI agents are still far from perfect. While AI-enhanced search tools are more reliable than general AI models, they are not immune to inaccuracies. GPT-4o search, for instance, still gets 10% of factual queries wrong, and citation reliability remains an issue. Moreover, the CUA model is still in early development and not yet fully reliable for automating tasks across operating systems.
To address these limitations, OpenAI is also launching the Agents SDK, an open-source toolkit that allows developers to integrate AI models into their systems, implement safety measures, and monitor agent activities for debugging and optimization. This follows OpenAI’s Swarm framework, released last year, which focused on multi-agent coordination.
The Future of AI Agents
OpenAI’s API product head, Olivier Godement, acknowledged the challenges of scaling AI agents, noting that while it’s easy to demo an agent, creating one that people use regularly is far more difficult. However, he believes that AI agents represent one of the most impactful AI applications today.
His sentiment echoes OpenAI CEO Sam Altman’s statement earlier this year that 2025 will be the year AI agents enter the workforce. Whether that prediction materializes remains to be seen, but with the Responses API and its accompanying toolkit, OpenAI is clearly shifting its focus from AI demos to practical, scalable solutions.