OpenAI Launches New Tools for Businesses to Build AI Agents | TechCrunch

1. OpenAI's New Tool: Responses API

On March 10, 2025, OpenAI announced new tools to help enterprises and developers build AI agents. The tool, OpenAI's Responses API, enables AI agents to automatically perform tasks like web search, company file exploration, and website navigation. It replaces the previously offered Assistants API, which will be phased out by the first half of 2026.

"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."

Olivier Godement, OpenAI API Product Lead

As this quote suggests, there's a big difference between simply demonstrating an AI agent and making it truly useful. OpenAI aims to close this gap with the Responses API.

2. Key Features of the Responses API

The Responses API leverages OpenAI's latest AI models and frameworks to provide various capabilities:

Web Search: The Responses API uses OpenAI's GPT-4o search and GPT-4o mini search models to search the web, cite sources, and generate answers.
- GPT-4o search achieved 90% accuracy, and GPT-4o mini search achieved 88% accuracy. These are significantly higher than OpenAI's previous GPT-4.5 model at 63%.
- However, a 10% error rate remains, with weaknesses on short exploratory queries (e.g., "What's the Lakers score today").
File Search: Includes the ability to quickly search files within company databases. OpenAI explicitly stated this data won't be used for training.
Computer Use Automation (CUA Model): The CUA model used in Operator generates mouse and keyboard actions, automating tasks like data entry and app workflows.
- Enterprises also have the option to run this model on local systems.

"The CUA model is not yet highly reliable for automating tasks on operating systems."

OpenAI Blog

The CUA model isn't yet perfect for automating OS-level tasks and may make mistakes, but OpenAI says they're continuously improving it.

3. Responses API and Agents SDK

Alongside the Responses API, OpenAI released the Agents SDK, an open-source toolkit that provides:

Integration of AI models with internal systems
Safety guard configuration
Monitoring and debugging AI agent activities

The Agents SDK can be seen as the successor to Swarm (a multi-agent coordination framework) that OpenAI announced last year. It enables developers to build more stable and optimized AI agents.

4. The Present and Future of AI Agents

AI agents have drawn significant attention in recent years, but many technical limitations and challenges remain. For example, Chinese startup Butterfly Effect's Manus AI agent platform generated buzz but disappointed by failing to deliver on many promises.

OpenAI aims to transition "from demos to real products" through tools like the Responses API.

"Agents are the most impactful application of AI that will happen."

Olivier Godement

OpenAI CEO Sam Altman has also projected 2025 as "the year AI agents become mainstream in the workplace," emphasizing that AI agents will be a core technology of the future.

5. Limitations and Room for Improvement

The Responses API and related tools open many possibilities, but challenges remain:

AI hallucination: The problem of AI generating incorrect information hasn't been fully resolved.
Source reliability: ChatGPT's cited sources aren't always trustworthy and need improvement.
CUA model reliability: Lack of reliability in OS task automation is a challenge to address.

OpenAI says these tools are in early stages and will deliver better results through continuous improvement.

6. Conclusion: OpenAI's Direction

OpenAI aims to elevate AI agent technology to the next level through the Responses API and Agents SDK. Going beyond flashy demos, their effort to provide truly useful and reliable tools stands out. Whether 2025 truly becomes "the year of the AI agent" remains to be seen, but OpenAI's latest announcement has certainly raised the bar.

"Whether or not 2025 truly becomes the 'year of the AI agent,' OpenAI's latest releases show the company wants to shift from flashy agent demos to impactful tools."

Maxwell Zeff, TechCrunch reporter