OpenAI's New Tools for Building AI Agents

OpenAI has just rolled out a set of new tools aimed at helping developers and businesses create AI agents—automated systems capable of performing tasks on their own. These tools are part of OpenAI’s latest Responses API, which allows businesses to develop custom AI agents that can perform tasks like web searches, sifting through company files, and navigating websites, similar to OpenAI’s Operator product. This new API is set to replace the existing Assistants API, which will be phased out by mid-2026.

The Growing Interest in AI Agents

The buzz around AI agents has been on the rise, although the tech industry is still grappling with clearly defining what these agents are. Recently, a Chinese startup called Butterfly Effect gained attention with its AI agent platform, Manus, which didn't quite meet user expectations. This highlights the pressure on OpenAI to deliver AI agents that truly work.

Olivier Godement, who leads OpenAI’s API product, mentioned that while it's easy to demonstrate an AI agent, scaling it and ensuring frequent usage is quite challenging. Earlier this year, OpenAI introduced two AI agents in ChatGPT: Operator, which helps with website navigation, and deep research, which compiles research reports. While these tools showed potential, they still lacked full autonomy.

What the Responses API Offers

With the new Responses API, OpenAI aims to provide access to the building blocks of AI agents, enabling developers to create applications similar to Operator and deep research but with enhanced autonomy. Developers can utilize the same AI models that power OpenAI’s ChatGPT Search tool: GPT-4o search and GPT-4o mini search. These models are designed to find answers online, citing sources as they go.

OpenAI boasts that these models are highly accurate, with GPT-4o search scoring 90% and GPT-4o mini search scoring 88% on their SimpleQA benchmark, which tests the models' ability to answer short, factual questions. For context, the larger GPT-4.5 model scores 63%.

The Responses API also includes a file search feature that can quickly comb through a company’s databases for information. OpenAI assures that it won't use these files to train its models. Additionally, developers can use OpenAI’s Computer-Using Agent (CUA) model, which powers Operator, to automate tasks like data entry and app workflows. This model can be run locally on enterprise systems, although the consumer version is limited to web actions.

Challenges and Future Improvements

Despite these advancements, the Responses API isn't a cure-all for the current challenges facing AI agents. AI-powered search tools, while accurate, can still produce errors—GPT-4o search gets 10% of factual questions wrong. Moreover, these tools sometimes struggle with brief, navigational queries, and their citations aren't always dependable.

OpenAI acknowledges that these tools are in their early stages and is committed to ongoing improvements. Alongside the Responses API, OpenAI is releasing an open-source toolkit called the Agents SDK. This toolkit provides developers with resources to integrate models with internal systems, establish safeguards, and monitor AI agent activities for optimization.

Godement expressed hope that OpenAI can bridge the gap between AI agent demonstrations and practical applications this year, emphasizing that agents could be one of AI's most significant applications. OpenAI CEO Sam Altman also projected that AI agents would become mainstream in the workforce by 2025.

Whether or not 2025 will indeed be the year of the AI agent, OpenAI’s latest developments indicate a shift from flashy demos to tools that aim to make a real impact.