OpenAI Unveils Responses API for Building AI Agents
Amid the buzz created by Chinese startups and their AI models, OpenAI recently unveiled the Responses API. It is a new tool designed to help developers and businesses create custom AI agents.
These AI agents, which can independently complete tasks like web searches, document scanning, and website navigation, are powered by OpenAI’s models. This release marks a significant step in OpenAI’s push to enhance its AI agent offerings and improve their autonomy. The release marks a step in OpenAI's push to enhance its AI agent offerings and improve their autonomy amid the growing competition.
Revolutionizing AI Agents for Businesses
The new Responses API allows businesses to create more powerful AI agents using OpenAI’s models, such as GPT-4o search and GPT-4o mini search. These tools can search the web, gather information, and generate responses with a high degree of factual accuracy.
Addressing the questions regarding its accuracy, OpenAI claims that GPT-4o search achieves 90% accuracy on the SimpleQA benchmark. This score outperfoms the ones set by GPT-4o mini search and GPT-4.5. They were 88% and 63%, respectively.
The API is also designed to help developers build AI agents that can automate tasks beyond than just browsing the web. For instance, the Responses API includes a file search utility. This allows businesses to scan internal databases for information. Moreover, the API provides access to OpenAI's Computer Using Agent (CUA) which automates tasks like data entry and app workflows.
For example, the Responses API includes a file search utility, allowing businesses to scan internal databases for information. Additionally, the API provides access to OpenAI’s Computer-Using Agent (CUA), which automates tasks like data entry and app workflows. Businesses can also run the CUA model locally on their own systems for increased security.
Addressing AI Agent Challenges
Despite the potential, AI agents face significant challenges. OpenAI acknowledges that web search tools, while accurate, still suffer from occasional errors and limitations in providing up-to-date answers. Additionally, OpenAI noted in a blog post provided to TechCrunch that the CUA model is "not yet highly reliable for automating tasks on operating systems" and is prone to making "inadvertent" mistakes. Furthermore, as Olivier Godement, OpenAI’s API product head, told TechCrunch in an interview, "It's pretty easy to demo your agent. To scale an agent is pretty hard, and to get people to use it often is very hard."
OpenAI is also releasing the Agents SDK, an open-source toolkit that helps developers integrate AI models with internal systems and monitor agent activities for optimization and debugging.
Looking Ahead
OpenAI’s goal is to bridge the gap between AI agent demos and functional tools that can be widely adopted. The company believes that AI agents will become the most impactful application of AI, with OpenAI CEO Sam Altman predicting that 2025 will be the year AI agents enter the workforce. As these tools evolve, OpenAI aims to shift from flashy demos to real-world solutions for businesses.
