OpenAI is developing AI agents: ChatGPT learns to perform work tasks independently

What are OpenAI AI Agents

Unlike a standard chatbot, an AI agent can independently execute a sequence of actions to achieve a set goal. The user formulates a task, and the system determines which tools and data it needs, then performs the work.

OpenAI is developing this approach in ChatGPT Work. The service is primarily aimed at office workers who deal with a large number of digital tools and information.

An AI agent can interact with email, Slack, Notion, Figma, Google Calendar, and other services. For example, a user can ask it to gather information from work messages, prepare a report, create a spreadsheet, or transfer data from one service to another.

OpenAI Wants to Take AI Agents Beyond Programming

Until now, one of the main areas of development for AI agents has remained software development. Tools like Codex and Claude Code already allow programmers to assign multi-step tasks to artificial intelligence.

However, OpenAI aims to extend this approach to other professions. The company considers accountants, investors, marketers, managers, and other office specialists as potential users of AI agents.

Within OpenAI itself, such tools are already used to prepare weekly reports, work with metrics, and create planning spreadsheets. Investors use AI agents to gather information about companies and prepare investment materials, while operations teams use them to create dashboards and visualize data.

Why AI Agents Need Access to Work Data

To perform tasks independently, an AI agent needs more than just a language model. It requires access to the information and tools that a human uses.

This is why OpenAI is developing what is called a harness—a software shell that determines what data the model receives, which tools it can use, and how it performs tasks.

For the average user, this technology should remain virtually invisible. The idea is that it is sufficient for a person to describe the task in plain language, and the system will figure out the rest of the actions on its own.

At the same time, this approach requires a large number of permissions. The user effectively grants AI access to part of their digital life, including work documents, messages, and calendars.

The Main Problem is Trust

The more capabilities an AI agent gains, the higher the potential risks.

If the system has access to email, work chats, and documents, it may obtain information that the user did not intend to use for a specific task. For example, when preparing a document, the model could theoretically access private correspondence.

Therefore, one of the key questions for the development of AI agents is balancing convenience with user control.

In practice, configuring permissions remains one of the less straightforward parts of working with such tools. Some features require full access to the service, and certain settings are only available through the web version.

OpenAI Competes with Anthropic

The development of AI agents is taking place against the backdrop of competition between the largest AI companies.

Anthropic previously bet on Claude Code, which became one of the most prominent tools for developers. OpenAI, in turn, is developing Codex and ChatGPT Work.

The companies' approaches differ primarily in the level of user interaction. Claude Code often offers several action options and regularly requests confirmation. OpenAI strives to make the process more automatic, so that the user can simply set a task and receive a ready result.

At the same time, OpenAI considers its own language models to be one of the main advantages, allowing agents to perform more complex tasks.

AI Agents Could Change Office Work

The potential of AI agents lies not only in automating individual operations. In the future, they could become digital assistants that independently work with information and perform tasks across multiple services.

For example, an employee could ask AI to collect data from corporate email and CRM, analyze it, create a report, and prepare a presentation. Currently, a person has to manually switch between several programs to do this.

However, the mass adoption of AI agents depends on several factors: model quality, computational costs, data security, and ease of use.

According to a study supported by OpenAI, 98% of company employees used Codex in June 2026. Among OpenAI's organizational subscribers, this figure was 17%, and among individual users, it was less than 1%. This disparity shows that the technology is currently much more popular within AI companies than among mass users. 

How Much Do AI Agents Cost

Another question relates to the cost of using such systems. AI agents perform significantly more operations than a standard chatbot, meaning they consume more tokens and computational resources.

With active use of ChatGPT Work, tens of millions of tokens can be spent in just a few days. This creates additional load on infrastructure for AI companies and makes model efficiency an important factor in the development of agentic systems.

OpenAI is already reducing the cost of using its models and expects that new generations of AI will be able to perform similar tasks with lower computational costs.

What's Next

OpenAI is gradually transforming ChatGPT from a tool for answering questions into a system capable of independently executing workflows. Previously, the company also launched OpenAI Frontier—a platform for creating and managing AI agents in a corporate environment.

The main question now is not just how intelligent the AI model will be. Its access to data, integrations with work services, security, and ability to understand the context of a task are also crucial.

If OpenAI succeeds in making work with AI agents simple and secure enough, they could become the next stage in the development of corporate artificial intelligence. Instead of individual chatbots, companies will get digital assistants capable of independently performing multi-step tasks and working with multiple tools simultaneously.