OpenAI Launches Tools for Building Corporate AI Agents

OpenAI is attempting to assist developers in creating their own agents, which are regarded as the AI of the future. Using its own AI models and frameworks, OpenAI on Tuesday unveiled new tools to assist developers and businesses in creating AI agents, or autonomous systems that can do tasks on their own. The business is launching a new Responses API that provides developers with building blocks to construct agents that can browse the web, find files, and carry out operations on a computer on their behalf.

The tools are a component of OpenAI’s new Responses API, which enables companies to create bespoke AI agents that, like OpenAI’s Operator product, can navigate websites, do online searches, and search through company data. The Responses API basically replaces OpenAI’s Assistants API, which the firm aims to discontinue in the first half of 2026.

Even though the computer sector has had difficulty defining or even demonstrating what “AI agents” actually are, the excitement around them has increased significantly in recent years. A Chinese firm named Butterfly Effect went viral earlier this week for a new AI agent platform called Manus, which consumers soon found didn’t live up to many of the company’s claims. This is the latest instance of agent hype outpacing usefulness.

To put it another way, OpenAI has a lot on the line to get agents right.

In an interview with TechCrunch, Olivier Godement, the leader of OpenAI’s API product, stated, “Demonstrating your agent is fairly simple.” “It’s difficult to scale an agent and to get people to use it frequently.” Oliver tells The Verge, “We will be able to build ourselves some agents, like Deep Research and Operator.” However, there are a lot of businesses and application cases, and the world is really complicated. Therefore, we are very happy to offer developers the foundations and building blocks they need to create the finest agents for their use cases and requirements.

Two AI agents were added to ChatGPT earlier this year by OpenAI: deep research, which creates research reports for you, and operator, who navigates websites on your behalf. Both tools demonstrated the potential of agentic technology, but they fell far short in the “autonomy” area.

In order to enable developers to create their own Operator- and deep research-style agentic apps, OpenAI now hopes to offer access to the parts that drive AI agents through the Responses API. With its agent technology, OpenAI intends to let developers design apps that seem more independent than those now on the market.

The same AI models (in preview) that power OpenAI’s ChatGPT Search online search tool—GPT-4o search and GPT-4o mini search—can be accessed by developers via the Responses API. The models are able to search the internet for answers to queries, referencing sources while they do so. Also with the help of the online search engine included in the Responses API, which is based on the same paradigm as ChatGPT, developers may utilize GPT-4o and GPT-4o small to retrieve real-time information and citations from the web. It also contains a computer-use capability that, in a similar manner, employs the Operator model of the organization to carry out tasks for a user. Additionally, the Responses API offers a tool for searching through vast amounts of papers; OpenAI bills it as a method to aid customer service representatives in sorting through frequently asked questions or as a way for a legal assistant to find prior instances.

GPT-4o search and GPT-4o mini search are very factually accurate, according to OpenAI. GPT-4o search receives a score of 90% and GPT-4o small search receives an 88% on the company’s SimpleQA benchmark, which evaluates models’ capacity to respond to brief, fact-seeking queries (the higher the better). In contrast, OpenAI’s considerably bigger, newly published model, GPT-4.5, only receives a score of 63%.

A file search tool that may swiftly browse through files in a business’s databases to get information is also included in the Responses API. (OpenAI asserts that these files will not be used to train models.) Additionally, OpenAI’s Computer-Using Agent (CUA) architecture, which underpins Operator, is accessible to developers via the Responses API. Data entry and app workflows may be automated by developers because of the model’s ability to produce mouse and keyboard movements.

According to OpenAI, businesses can choose to operate the CUA model locally on their own systems when it launches in research preview. The Operator’s consumer version of the CUA is limited to online action.

To be clear, not every technical issue facing AI agents today will be resolved by the Responses API.

Although it is not unexpected that AI-powered search tools are more accurate than standard AI models since they can simply look up the correct answer, online search does not solve the issue of AI hallucinations. Ten percent of factual queries are still incorrectly answered by GPT-4o search. In addition to their accuracy, AI search engines sometimes have trouble answering brief, basic questions (such as “Lakers score today”), and new studies indicate that ChatGPT’s citations aren’t always trustworthy.

OpenAI stated in a blog post given to TechCrunch that the CUA model is prone to “inadvertent” errors and is “not yet highly reliable for automating tasks on operating systems.”

These are preliminary versions of OpenAI’s agent tools, though, and the company is always trying to make them better.

The Agents SDK, an open-source toolkit that OpenAI is releasing alongside the Responses API, gives developers free tools to link models with their internal systems, implement security measures, and track AI agent activity for debugging and optimization. A sort of sequel to OpenAI’s Swarm, a multi-agent orchestration framework that the firm unveiled late last year, is the Agents SDK.

According to Godement, “agents are the most impactful application of AI that will happen,” and he thinks OpenAI can close the gap between AI agent demos and products this year. This is in line with OpenAI CEO Sam Altman’s declaration from January that 2025 will see the introduction of AI agents into the workforce.

The release of the Agents SDK and Responses API expands on the tools that OpenAI currently provides developers, such as the Chat Completions API, which enables developers to create AI applications that provide answers to user inquiries. In the middle of 2026, OpenAI also intends to switch from the Assistants API to the Responses API. The business claims that by utilizing the input developers gave on the Assistants API, it “incorporated key improvements into the Responses API.”

OpenAI’s website has further information about its new developer tools.

Regardless of whether 2025 ends up being the “year of the AI agent,” OpenAI’s most recent releases indicate the firm aims to move away from dazzling agent demonstrations and toward useful tools.