OpenAI Launches New AI Agent Tools for Businesses

OpenAI Launches New Tools for Building AI Agents

On Tuesday, OpenAI unveiled a suite of new tools specifically engineered to empower developers and enterprises in the creation of AI agents. These automated systems are designed to independently execute tasks, leveraging OpenAI’s proprietary AI models and frameworks.

Introducing the Responses API

These new capabilities are centered around OpenAI’s Responses API. This API enables businesses to develop customized AI agents capable of performing web searches, accessing company files, and navigating websites – functionalities similar to those found in OpenAI’s Operator product. Notably, the Responses API is intended to supersede OpenAI’s Assistants API, with the latter slated for discontinuation in the first half of 2026.

The Hype and Reality of AI Agents

The enthusiasm surrounding AI agents has surged in recent years. However, the technology industry has faced challenges in demonstrating practical applications and even establishing a clear definition of what constitutes an “AI agent.” Recent instances, such as the viral launch of Chinese startup Butterfly Effect’s Manus platform, highlight the gap between promise and performance.

Therefore, successful implementation is crucial for OpenAI.

Scaling Agent Technology

“Demonstrating an agent’s capabilities is relatively straightforward,” explained Olivier Godement, API product head at OpenAI, in a TechCrunch interview. “However, achieving scalability and consistent user engagement presents significant hurdles.”

Early Agent Precursors: Operator and Deep Research

Earlier this year, OpenAI introduced two AI agents within ChatGPT: Operator, which automates website navigation, and a deep research tool for compiling reports. While these tools showcased the potential of agentic technology, their level of autonomy remained limited.

The Responses API aims to provide developers with access to the underlying components powering these agents. This will allow them to build applications with similar Operator and deep research functionalities, with the goal of achieving greater autonomy.

Leveraging GPT-4o Search and GPT-4o Mini Search

Developers utilizing the Responses API can now access the same AI models that drive OpenAI’s ChatGPT Search tool: GPT-4o search and GPT-4o mini search. These models can conduct web searches to answer questions, providing source citations alongside their responses.

Accuracy and Performance Metrics

OpenAI asserts that GPT-4o search and GPT-4o mini search exhibit a high degree of factual accuracy. On the company’s SimpleQA benchmark – measuring the ability to answer fact-based questions – GPT-4o search achieved a score of 90%, while GPT-4o mini search scored 88%. For context, OpenAI’s larger GPT-4.5 model scored 63% on the same benchmark.

File Search and the Computer-Using Agent

The Responses API also incorporates a file search function, enabling rapid scanning of company databases for information retrieval. OpenAI assures users that models will not be trained on these files. Furthermore, developers can utilize OpenAI’s Computer-Using Agent (CUA) model, which powers Operator, to generate mouse and keyboard actions for automating computer tasks like data entry and workflow management.

Enterprises have the option to run the CUA model locally on their own infrastructure, currently available in research preview. The consumer version of CUA, as seen in Operator, is limited to web-based actions.

Acknowledging Existing Limitations

It’s important to note that the Responses API does not represent a complete solution to the challenges facing AI agents today.

While AI-powered search tools demonstrate improved accuracy compared to traditional AI models, they do not entirely eliminate the risk of AI hallucinations. GPT-4o search still produces incorrect answers 10% of the time. Additionally, these tools can struggle with concise, navigational queries and concerns regarding citation reliability have been raised.

OpenAI acknowledges that the CUA model is “not yet highly reliable” for automating operating system tasks and may produce unintended errors.

Ongoing Development and Improvement

OpenAI emphasizes that these are early iterations of their agent tools and that continuous improvement is a priority.

The Agents SDK: An Open-Source Toolkit

Alongside the Responses API, OpenAI is releasing the Agents SDK, an open-source toolkit. This toolkit provides developers with free tools to integrate models with internal systems, implement safety measures, and monitor AI agent activities for debugging and optimization.

The Agents SDK builds upon OpenAI’s Swarm framework, released late last year, for multi-agent orchestration.

The Future of AI Agents

Godement expressed optimism that OpenAI can bridge the gap between agent demonstrations and practical products this year. He believes that “agents are the most impactful application of AI that will happen,” echoing a statement made by OpenAI CEO Sam Altman, who predicted that 2025 will be the year AI agents enter the workforce.

Regardless of whether 2025 marks the “year of the AI agent,” OpenAI’s recent releases demonstrate a commitment to transitioning from showcasing agent potential to delivering impactful tools.