How much control are you willing to give an AI model over your digital life? For OpenAI, that question has moved from hypothetical to product strategy. ChatGPT Work, released last month on the company's $20-per-month subscription tier, lets non-engineers put AI agents to work in their inbox, Slack workspace, Notion documents, Figma files, and other everyday tools. The model doesn't just answer questions; it can complete multistep projects, reach into connected services, and act on the information it finds.

Andrew Ambrosino, the lead engineer for OpenAI's desktop app, is the product's most exposed test user. He has given the app access to his inbox, Slack, phone, and project management applications. “If I'm asking it to write a document, is there a possibility that it's going to pull from a private DM on that subject and not know that it's not supposed to share some info? Yes,” he said. Still, he says the risk is part of the job. “I will take the personal hit here and there if I have to. And I haven't had to.”

The launch comes as OpenAI and its rivals search for the next wave of AI growth. Coding tools proved that LLMs could do more than generate text, but programmers are a small slice of the workforce. If AI labs are to justify massive spending on models and data centers, they need to reach lawyers, accountants, investors, doctors, and managers. That means making agents that can work with the messy, human-centered tools that dominate office life.

From coding assistant to general-purpose agent

ChatGPT Work is a modified version of Codex, OpenAI's coding tool. Codex showed that an AI model could edit code, run tests, and manage software projects. But the company believes the same technology can handle marketing reports, financial memos, and operations dashboards. Thibault Sottiaux, who leads OpenAI's core product work, says the goal is to let ChatGPT complete “entire, very complicated tasks for you all autonomously in a way that is delightful and safe.”

Commercially, that matters. Longer agentic tasks consume more tokens, which makes them more lucrative per user. Reaching new professions is crucial not just for OpenAI but for the entire industry. Vertical-specific competitors like Harvey, which targets lawyers, and Clay, which targets sales teams, are already chasing those customers with a model-agnostic approach. They can plug whichever AI model works best at the time.

The gap between internal and external adoption shows the size of the challenge. An OpenAI-backed study found that in June, 98% of OpenAI employees were using Codex, but just 17% of organizational subscribers and less than 1% of individual subscribers were using the agentic coding tool. That difference between near-total adoption inside the company and negligible adoption outside it defines the opportunity for ChatGPT Work.

Making the harness less intimidating

Every LLM runs inside a “harness,” the software that decides what information the model sees, which tools it can use, and how it presents results. For developers, a command-line interface was enough to change software development. But most people don't use command lines. “There's a reason Windows replaced DOS,” Ambrosino notes.

OpenAI's engineers describe ChatGPT Work as an attempt to bring the power of coding agents to a mass audience. The product includes more buttons for selecting projects and plug-ins, but the core experience is still a chat window. Ambrosino compares the design to skeuomorphism, a fading practice in which digital tools imitated physical objects. “That stuff wasn't just cringe design. That actually helped get people into this [and] make the transition.”

One OpenAI engineer described asking the program to review a Slack conversation about an engineering problem and “make some charts,” then receiving back a series of insightful plots. Akshay Nathan, who leads the product engineering team, says the tool helps with the “deluge of information” facing knowledge workers. “We're actually quite limited by our ability to parse everything that's available to us, and then take action on it,” he said.

OpenAI won't disclose exact user numbers for Work versus Codex, but says the combined app is used by 20 million people, compared with more than a billion users who prompt ChatGPT online.

Testing the agent in the real world

In practice, the system can be impressive. One journalist testing the tool asked it to extract a child's oddly formatted preschool calendar from an inbox and transfer it to Google Calendar, saving much manual data entry. The same tool produced a real-time financial dashboard for publicly traded companies and a queryable database of space launches, tasks that had previously required writing Python scripts.

But friction remains. Setting up permissions for cloud drive access can be confusing and circular. Many important settings are only available in the web app, forcing users to switch between mobile and desktop. The model can create calendar events but not new calendars. And if users don't set the “effort” level high enough, the output can feel like the worst intern they've ever worked with.

Joe Gershenson, engineering lead for OpenAI's harness, admits the effort setting is not intuitive yet. “There are things that we can do better to help them get the right level of reasoning,” he said.

Learning from Claude Code

OpenAI didn't invent the agentic harness. Anthropic's Claude Code and Claude Cowork defined much of the current product category. The two companies are now competing directly for the same users.

OpenAI engineers are reluctant to discuss competitors. “The Mad Men 'I don't think about you at all' meme comes to mind here,” Gershenson said. Yet the products' interfaces look strikingly similar, and the first time ChatGPT Work is launched, it prompts users to import data from Claude Cowork.

History explains some of the tension. OpenAI initially developed Codex as a web app with an ambitious, fully autonomous approach. It bet the model could handle entire tasks on its own with minimal input. Anthropic's Claude Code took a more conversational path: it would survey possibilities, offer options, check back, and keep the user in the loop.

That second approach proved more effective. “Our product was a little ahead of where the model and harness was at the time,” Ambrosino says. OpenAI later added more checkpoints and interactions, creating the Codex that exists today.

Is the harness the moat?

There is an ongoing debate about whether a great harness matters as much as a great model. Some AI researchers point to the “bitter lesson”: better general models beat hand-crafted domain-specific features. Gershenson argues that the next model often makes elaborate harness features obsolete. His team tries to expose exactly the information and tools the model needs, and nothing more.

Independent tests suggest the answer isn't simple. Comparisons by Composio and Databricks show that different harness and model combinations produce different results. Dat

Source: TechCrunch News