OpenAI Introduces New Tools for Businesses to Develop AI Agents
On Tuesday, OpenAI unveiled new tools aimed at helping developers and businesses create AI agents—automated systems capable of independently performing tasks—using the company’s AI models and frameworks.
These tools are part of OpenAI’s new Responses API, which enables businesses to develop custom AI agents that can conduct web searches, scan internal files, and navigate websites, similar to OpenAI’s Operator product. The Responses API replaces the Assistants API, which OpenAI plans to phase out by mid-2026.
The Hurdles of AI Autonomy
Despite growing excitement around AI agents, the industry has struggled to clearly define or demonstrate their practical value. A recent example is Chinese startup Butterfly Effect’s Manus platform, which went viral but failed to meet many user expectations, highlighting the challenges of delivering truly autonomous AI.
OpenAI aims to overcome these hurdles. “It’s pretty easy to demo your agent,” said Olivier Godemont, OpenAI’s API product head, in an interview with TechCrunch. “To scale an agent is pretty hard, and to get people to use it often is very hard.”
Earlier this year, OpenAI introduced two AI agents in ChatGPT: Operator, which navigates websites, and Deep Research, which compiles research reports. While these tools showcased agentic capabilities, they lacked full autonomy.
With the Responses API, OpenAI now offers businesses access to the core components behind its AI agents, allowing developers to build their own applications that could surpass current solutions in autonomy and usability.
With the Responses API, developers can access the same AI models powering OpenAI’s ChatGPT Search tool: GPT-4o search and GPT-4o mini search. These models can browse the web for answers, citing sources as they generate responses.
OpenAI claims these models are highly accurate. On its SimpleQA benchmark, which evaluates fact-based question answering, GPT-4o search scores 90%, while GPT-4o mini search scores 88%—outperforming the recently released GPT-4.5, which scores only 63%.
Limitations of AI-Powered Search
AI-powered search tools generally surpass traditional AI models in accuracy since they can look up information directly. However, they still struggle with certain challenges, including hallucinations and difficulties with short, navigational queries like “Lakers score today.” Reports also suggest that ChatGPT’s citations are not always reliable.
The Responses API also features a file search utility that quickly retrieves information from a company’s databases. OpenAI assures that these files won’t be used for model training. Additionally, developers can integrate OpenAI’s Computer-Using Agent (CUA) model, which powers the Operator tool. This model generates mouse and keyboard actions, enabling automation for tasks like data entry and workflow management.
Enterprises can choose to run the CUA model locally on their systems, as it is launching in a research preview. However, the consumer version available in Operator is limited to web-based actions.
Despite these advancements, the Responses API doesn’t eliminate all technical hurdles in AI agents. GPT-4o search still provides incorrect answers 10% of the time, and OpenAI acknowledges that its CUA model is not yet fully reliable for automating operating system tasks, as it can make unintended errors.
To support developers, OpenAI is also launching the Agents SDK, an open-source toolkit that helps integrate AI models with internal systems, implement safeguards, and monitor agent behavior for debugging and optimization. The SDK builds on OpenAI’s Swarm framework, released last year for multi-agent orchestration.
OpenAI’s API product lead, Olivier Godemont, believes this year will be crucial in turning AI agents from demos into practical tools. CEO Sam Altman has similarly predicted that 2025 will be the year AI agents enter the workforce. Whether that vision materializes remains to be seen, but OpenAI’s latest releases signal a shift toward making AI agents more functional and impactful.
Read the original article on: TechCrunch
Leave a Reply