What Are the New OpenAI Tools Launched to Help AI Agent Development?

2025-03-17

OpenAI has unveiled a suite of powerful new tools designed to bridge the gap between flashy AI agent demos and practical, production-ready applications. These tools aim to simplify the development of AI agents—automated systems capable of independently accomplishing tasks on behalf of users—and address the significant challenges developers face when building agentic applications.

OpenAI introduced new tools for building AI agents. Image credit: OpenAI

The Evolution of AI Agents and Their Challenges

AI agents represent what many industry experts believe could be the most transformative application of artificial intelligence. OpenAI’s API Product Head, Olivier Godement, has gone so far as to call them “the most impactful application of AI that will happen,” echoing CEO Sam Altman’s prediction that 2025 would be the year AI agents enter the workforce.

Despite the growing enthusiasm, turning advanced AI models into reliable agents has proven difficult. Developers have struggled with:

Extensive prompt iteration
Complex custom orchestration logic
Limited visibility into agent behavior
Insufficient built-in support for agent development

The gap between demos and practical applications has been particularly challenging to overcome. While it may be “pretty easy to demo your agent,” as Godement noted, “to scale an agent is pretty hard, and to get people to use it often is very hard.”

The Responses API: A New Foundation for Agent Development

At the heart of OpenAI’s new offerings is the Responses API, which effectively replaces the Assistants API (set to sunset in mid-2026). The Responses API combines the simplicity of the Chat Completions API with the tool-use capabilities of the Assistants API, creating a more flexible foundation for building agentic applications.

With a single API call, developers can now:

Solve increasingly complex tasks using multiple tools
Handle multiple model turns in one request
Leverage built-in tools that connect models to the real world
Benefit from unified item-based design and intuitive streaming events

The API provides developers with everything needed to easily combine OpenAI models and built-in tools without the complexity of integrating multiple APIs or external vendors.

Built-in Tools for Practical Agent Capabilities

Web Search: Real-time Information with Citations

Developers can now integrate up-to-date web search capabilities powered by the same models that drive ChatGPT search: GPT-4o search and GPT-4o mini search. These models have demonstrated impressive factual accuracy, scoring 90% and 88% respectively on the SimpleQA benchmark—significantly outperforming the larger GPT-4.5 model’s 63%.

Early adopters like Hebbia have already implemented web search to help asset managers, private equity firms, and law practices extract actionable insights from extensive datasets. Their applications deliver context-specific market intelligence that outperforms current benchmarks.

Importantly, responses generated with web search include clear, inline citations to sources, giving users a way to verify information while providing content owners with new opportunities to reach broader audiences.

File Search: Efficient Document Retrieval

The improved file search tool allows developers to easily retrieve relevant information from large volumes of documents. Key features include:

Support for multiple file types
Query optimization
Metadata filtering
Custom reranking
Fast, accurate search results

Companies like Navan have implemented file search in their AI-powered travel agent to quickly provide users with precise answers from knowledge-base articles. The tool enables them to set up powerful retrieval-augmented generation (RAG) pipelines without extra tuning or configuration.

Computer Use: Automating Tasks Across Systems

Perhaps most intriguing is the computer use tool, powered by the same Computer-Using Agent (CUA) model that enables OpenAI’s Operator feature. This research preview model has set new state-of-the-art records for computer use tasks:

38.1% success on OSWorld for full computer use tasks
58.1% success on WebArena
87% success on WebVoyager for web-based interactions

The tool captures mouse and keyboard actions generated by the model, allowing developers to automate computer tasks by translating these actions into executable commands within their environments.

Unify and Luminai have already integrated the computer use tool to automate complex workflows that were previously impossible to address with traditional APIs. Luminai, for instance, automated application processing and user enrollment for a major community service organization in days—a task that traditional robotic process automation (RPA) struggled with for months.

The Agents SDK: Orchestrating Agentic Workflows

Beyond the core API, OpenAI has released an open-source Agents SDK that simplifies orchestrating multi-agent workflows. This SDK improves upon their experimental Swarm framework released last year, with significant enhancements:

Easily configurable agents: LLMs with clear instructions and built-in tools
Intelligent handoffs: Seamless transfer of control between different agents
Configurable guardrails: Safety checks for input and output validation
Comprehensive tracing: Visualization tools to debug and optimize performance

Companies like Coinbase have used the Agents SDK to quickly prototype and deploy AgentKit, enabling AI agents to interact with crypto wallets and on-chain activities. Similarly, Box created agents leveraging web search and the SDK to help enterprises search, query, and extract insights from unstructured data across both internal and public sources.

Real-World Applications and Future Potential

These tools enable the development of practical AI agents for numerous use cases:

Customer support automation: AI agents that can access FAQs and provide accurate responses
Research assistants: Systems that compile research from multiple sources
Workflow automation: AI agents that navigate legacy systems without API availability
Shopping assistants: AI that can search the web and provide personalized recommendations
Legal and financial research: Tools that quickly reference past cases or market data

While these tools represent significant progress, OpenAI acknowledges certain limitations. Web search tools still get approximately 10% of factual questions wrong, and the Computer-Using AI Agent is “not yet highly reliable for automating tasks on operating systems,” with a tendency to make inadvertent mistakes.

Building the Platform for AI’s Future

OpenAI’s strategic shift from flashy demos to practical tools signals their belief that AI agents will soon become integral to the workforce. As model capabilities become increasingly agentic, the company plans to continue investing in deeper integrations across their APIs and new tools to help deploy, evaluate, and optimize agents in production.

Their goal is to provide developers with a seamless platform experience for building agents that can assist with various tasks across industries—addressing the critical gap between agent hype and practical utility that has characterized the field thus far.

For developers interested in exploring these new capabilities, the Responses API is available to all developers today, with pricing based on standard token and tool rates. The computer use tool is available as a research preview for select developers in higher usage tiers.

As OpenAI continues to refine these tools and address remaining technical challenges, we may indeed see 2025 become the year AI agents meaningfully enter the workforce, fulfilling the promise of autonomous systems that deliver real-world impact.

If you are interested in this topic, we suggest you check our articles:

Sources: OpenAI, TechCrunch

Written by Alius Noreika