OpenAI released an updated tool search API (Application Programming Interface, which allows different software to communicate) in March 2026. This update lets GPT agents (automated AI programs that act independently on tasks) find and load tool definitions as needed, saving tokens and enabling complex workflows to run more efficiently.
Building on this, GPT 5.4 allows agents to handle larger tool sets and to dynamically select the required tools without upfront loading.
Key Features and Capabilities
This tool search upgrade aligns with other recent responses and API enhancements, all aimed at building stronger, more reliable AI agents.
Main features include:
- Dynamic tool loading allows models to load large tool schemas only when needed, reducing initial token use and costs for tool-heavy apps, while supporting enhanced agentic workflows for multi-step tasks and improved accuracy through relevant tool selection and reasoning.
- Enhanced Agentic workflows. The system enables agents to plan, execute, and verify multi-step tasks more efficiently across different apps and data sources. Tools such as Web and File Searcher, and computer use enable them to access real-time information, analyze documents, and control computer interfaces.
- Better accuracy: by using relevant tools and improved reasoning. Agents select tools and complete tasks more accurately per internal benchmarks.
- Custom Tools Support: Developers can now make custom functions, namespaces (distinct naming systems for tools to avoid conflicts), or MCP (Multi-Channel Processing) servers searchable, allowing agents to connect with third-party services and private databases.
Consequently, the update moves the API toward more sophisticated, long-running agent applications. Agents now enjoy more independence and can serve as reliable digital workers.
In a related move, on Tuesday, OpenAI launched tools that let developers and businesses create AI agents automated programs powered by OpenAI’s models and frameworks to complete tasks independently.
OpenAI’s new Responses API lets businesses build custom AI agents to search the web, review files, and browse websites. It will replace the Assistant API, which is scheduled to retire in early 2026.
Interest in AI agents has increased in recent years, even though the tech industry still struggles to clearly define them or demonstrate their limitations. For example, earlier this week, the Chinese startup Butterfly Effect went viral for its new AI agent platform, Manus, but users soon found it did not deliver many of the company’s promised capabilities, highlighting new challenges with reliability and functionality.
These developments increase expectations for OpenAI to deliver AI agents that reliably solve industry challenges and perform in real-world environments.
It’s pretty easy to demo your agent, Oliver Godement, OpenAI’s product lead, told TechCrunch in an interview. Creating an agent is pretty hard, and getting people to use it is even harder.
Earlier this year, OpenAI added two AI agents to ChatGPT Operator, which browses websites for you, and Deep Search, which creates research reports. While these tools demonstrated agent technology, they did not achieve full independence and still require significant improvement in reliability and autonomy.
Now, with the responses of API, OpenAI is offering the building blocks behind its AI agents. Developers can use these to create their own apps similar to Operator and conduct deep research. Open AI works; this will lead to more independent-feeling AI applications than existing now. OpenAI hopes developers can tap into the same AI model (in preview) under the hood of its ChatGPT search tools, GPT-4O Search and GPT-4O Mini Search. The models can browse the web for answers to questions, citing sources as they generate replies.
OpenAI reports that GPT-4-0 search and GPT-4-0 mini search are highly accurate on OpenAI’s Simple QA test for factual question answering. GPT-4-0 search scored 90%, and GPT-4-0 mini search scored 88%, while GPT-4.5 scored 63%.
The responses API includes a file search utility that quickly scans company databases to retrieve information (OpenAI claims it won’t train models on these files). Developers can also use OpenAI’s Computer Using Agent (CUA), an AI that simulates mouse and keyboard actions, to automate tasks such as data entry and app workflows.
Enterprises can optionally run the CUA (Computer Using Agent) model, which is being released in Research Preview, locally on their own systems. OpenAI said: “The consumer version of the CUA available within Operator can only take actions on the web.”
It is important to note that the Responses API does not resolve all technical issues facing AI agents. While AI-powered search tools outperform traditional AI models in accuracy—unsurprisingly, as they can retrieve the correct answer, web search does not solve the hallucination problem. For example, GPT-4O Search still gets 10% of factual questions wrong.
Beyond accuracy, AI search teams also tend to struggle with short navigation queries, such as ‘Lakers’ scores today’. Recent reports suggest that ChatGPT’s citations aren’t always reliable.
OpenAI also acknowledges that the CU8 model is not yet reliable for automating operating system tasks and may make mistakes.
However, OpenAI says these agent tools are still in the early stages and that the company is continually working to improve them.
OpenAI is also launching an open-source agents SDK, giving developers tools to connect models, safeguards, and track agent activity. The SDK builds on OpenAI’s strong framework for managing multiple agents.
Godement said he hopes OpenAI can bridge the gap between AI agent demos and products this year, and that, in his opinion, agents will be the most impactful application of AI. This repeats a proclamation made by OpenAI CEO Sam Altman in January that 2025 is the year AI agents enter the workforce.
OpenAI’s newest tools show a clear shift: the company now prioritizes real usefulness and enterprise value over spectacle.
Source: OpenAI launches new tools to help businesses build AI agents
