On Tuesday, OpenAI unveiled a suite of new tools aimed at empowering developers and businesses to create AI agents—automated systems capable of performing tasks independently—leveraging the company’s own AI models and frameworks.
These tools fall under OpenAI’s newly launched Responses API, which allows enterprises to design custom AI agents that can conduct web searches, sift through internal documents, and browse websites, reminiscent of OpenAI’s existing Operator product. The Responses API is set to replace OpenAI’s Assistants API, which is due for deprecation in the first half of 2026.
The enthusiasm surrounding AI agents has surged in recent years, despite the technology sector’s challenges in defining or demonstrating what “AI agents” truly represent. A recent illustration of this phenomenon emerged when the Chinese startup Butterfly Effect went viral with its AI agent platform, Manus, which users quickly found to be underwhelming in fulfilling its promised capabilities.
This places significant pressure on OpenAI to develop effective AI agents.
“Demonstrating your agent is relatively simple,” Olivier Godemont, OpenAI’s head of API products, explained to TechCrunch in an interview. “However, scaling that agent and ensuring regular usage is considerably challenging.”
Earlier this year, OpenAI rolled out two AI agents within ChatGPT: Operator, which allows users to navigate websites automatically, and Deep Research, which assists in compiling research reports. While these tools showcased the potential of agentic technology, they left much to be desired in terms of autonomy.
With the introduction of the Responses API, OpenAI aims to provide developers with access to the foundational elements needed to create their own applications similar to Operator and Deep Research. The company hopes that this will inspire developers to produce applications with enhanced autonomy that surpass current offerings.
Through the Responses API, developers can leverage the same AI models powering OpenAI’s web search tool, ChatGPT Search: GPT-4o search and GPT-4o mini search. These models can seek answers online while providing source citations as responses are generated.
OpenAI asserts that GPT-4o search and GPT-4o mini search boast impressive factual accuracy. On the SimpleQA benchmark, which evaluates models on their capacity to answer concise fact-based questions, GPT-4o search achieves a score of 90%, while GPT-4o mini search scores 88%, both significantly above GPT-4.5, which garnered a score of just 63%.
While the accuracy of AI-powered search tools relative to traditional AI models may not be surprising—since GPT-4o search can simply look up correct information—it does not eliminate the issue of AI hallucinations. Furthermore, AI search tools continue to face difficulties with brief, navigational queries (like “Lakers score today”), and recent findings indicate that ChatGPT’s citations are not always dependable.
The Responses API also features a file search function that can quickly comb through a company’s data repositories to extract information (OpenAI states that these files will not be used to train models). Moreover, developers can access OpenAI’s Computer-Using Agent (CUA) model, which powers the Operator, enabling the automation of computer tasks such as data entry and application workflows through the generation of mouse and keyboard actions.
Organizations have the option to run the CUA model locally, which is in research preview, on their own systems, according to OpenAI. The consumer variant of the CUA available in Operator, however, is limited to web-based actions.
It’s important to note that the Responses API will not resolve all technical challenges currently faced by AI agents.
Despite the greater accuracy of AI-powered search tools over traditional models—a result of their ability to simply look up correct answers—web search does not fully address the issue of AI hallucinations. GPT-4o search still misinterprets 10% of factual inquiries. In addition to those inaccuracies, AI search tools often encounter obstacles with brief navigational questions (such as “Lakers score today”), with research suggesting that ChatGPT’s citations are not always trustworthy.
In a blog post shared with TechCrunch, OpenAI acknowledged that the CUA model is “not yet highly reliable for automating tasks on operating systems,” and is prone to making “inadvertent” errors.
Nonetheless, OpenAI emphasized that these are early-stage iterations of their agent tools, with ongoing efforts to refine and enhance them.
In tandem with the Responses API, OpenAI is launching an open-source toolkit called the Agents SDK, providing developers with free resources to integrate models into their internal systems, establish safeguards, and supervise AI agent activities for purposes of debugging and optimization. The Agents SDK serves as a successor to OpenAI’s Swarm, a framework for coordinating multiple agents that was released late last year.
Godemont expressed hope that OpenAI can bridge the divide between AI agent demonstrations and actual products this year, stating that in his view, “agents are the most transformative application of AI that will emerge.” This sentiment mirrors a statement made by OpenAI CEO Sam Altman in January, proclaiming that 2025 will be the year AI agents enter the labor market.
Whether or not 2025 becomes the definitive “year of the AI agent,” OpenAI’s recent offerings indicate a desire to transition from eye-catching agent demonstrations to substantial, impactful technologies.
Compiled by Techarena.au.
Fanpage: TechArena.au
Watch more about AI – Artificial Intelligence
