AI agents are no longer science fiction — they’ve landed in your daily life with a pretty direct question: how much control over your digital life would you be willing to hand over to an AI? 🤖
It might sound like a philosophical question, but in practice, it’s exactly what OpenAI is putting on the table with ChatGPT Work, its most ambitious product to date. The core idea is simple: instead of just answering questions, the AI starts executing entire tasks on its own, connecting emails, calendars, spreadsheets, project management tools, and much more. It’s no longer about typing a prompt and waiting for a response. It’s about delegating real responsibilities to a system that acts, decides, and executes.
And to prove this actually works, Andrew Ambrosino, the lead engineer for OpenAI’s desktop app, went way beyond any controlled experiment — he gave ChatGPT access to his inbox, Slack, Notion, Figma, and even his phone. If I ask it to write a document, is there a chance it pulls something from a private DM and doesn’t know it shouldn’t share that? Yes, Ambrosino told TechCrunch. I do this for the job. I’ll take the personal hit if I need to. And so far I haven’t needed to.
That’s the dilemma OpenAI is trying to solve while going all in on the next phase of artificial intelligence. Real risk, real reward.
What are AI agents and why they change everything
To understand the scale of this shift, it helps to take a step back and think about what AI agents actually represent within the LLM ecosystem. For a long time, Large Language Models were incredible text generation tools, but still passive. You asked, they answered. You requested, they delivered. The control stayed in your hands, and the AI functioned more like a super-advanced typing assistant than a truly autonomous collaborator.
AI agents flip that logic in a radical way. Instead of just processing an input and generating an output, they identify a goal, plan the steps needed to reach it, interact with external tools, make decisions along the way, and adjust course when something doesn’t go as expected. The model stops being reactive and becomes proactive within a defined scope.
That’s exactly the premise behind ChatGPT Work, launched last month and already available on the company’s most basic plan at 20 dollars a month. According to Thibault Sottiaux, who oversees core product work at OpenAI, including Work, the app can now accomplish entire and very complex tasks in a fully autonomous way, in a manner that’s delightful and safe. He sums up the company’s mission as bringing everyone along.
ChatGPT Work: OpenAI’s bet on the professional environment
ChatGPT Work isn’t an incremental update. In practice, it’s a modified version of Codex, the company’s own coding tool. The idea is to give non-engineers the same functionality that software developers already have with agents: an AI that doesn’t just answer questions but completes multi-step projects on its own.
Receive the best innovation content in your email.
All the news, tips, trends, and resources you're looking for, delivered to your inbox.
By subscribing to the newsletter, you agree to receive communications from Método Viral. We are committed to always protecting and respecting your privacy.
Commercially, this matters a lot. Agents that work for longer periods consume more tokens, making them more profitable for OpenAI on a per-user basis. Reaching new professions is crucial — not just for OpenAI, but for the entire industry. While programming has already proven to be profitable territory for AI labs, it’s still a small slice of the professional work these companies need to reach to justify their massive investments in training and compute. While the labs focused on software engineers, vertical competitors like Harvey in legal and Clay in sales went after those clients with a model-agnostic approach, plugging in whichever AI works best at any given moment.
The numbers show the size of the challenge. A study backed by OpenAI revealed that in June, 98% of the company’s employees used Codex, but only 17% of organizational subscribers and less than 1% of individual subscribers were using the tool. That gap between near-total adoption inside the company and negligible adoption outside it is both the challenge and the opportunity.
To make the app work for non-engineers, much more support was needed. According to Ambrosino, OpenAI’s own communications and finance teams started using Codex at a time when it was actively hostile to them, asking about code and showing things that made no sense. It was only between February and now that the tool gained a more general-purpose direction.
How to make AI intuitive
To understand that disconnect, it helps to understand what OpenAI engineers are building. Every LLM needs what engineers call a harness — the software that wraps around the model and decides what information it sees, which tools it can use, and how it presents responses. If you want the model to do things, to become an agent, the harness provides the tools and instructions for using them on long-running tasks.
For developers, a command-line interface was enough to change the way software is built. But most people don’t use that kind of interface — there’s a reason Windows replaced DOS. An agentic product that goes beyond software engineering, as Ambrosino explains, needs to deal with the messy world of your life, your tools, and websites that were built in 1995 and never updated.
He compares the effort to skeuomorphism, that now-outdated practice of making digital tools look like the physical objects they replaced, like a calculator app that looks like a pocket calculator. That wasn’t just cheesy design. That actually helped people make the transition, he says. Internally, there are debates: some employees argue that buttons are unnecessary if the user can simply ask the model. Ambrosino disagrees, at least for now: We push back on that, because it’s still too early. Discovery matters at this stage, and at some point we won’t have the button anymore. 🔍
Smart workflows: what actually changes in practice
When we talk about workflows transformed by AI agents, it’s worth thinking about concrete examples. For now, OpenAI positions the tool as ideal for routine, data-intensive tasks. Company employees are building weekly metrics reports and turning spreadsheets into planning tools.
Venture capitalists are using agents to gather communications and analyses about companies into investment memos, and operations teams are creating custom dashboards. Even Sam Altman uses the tool to plan his vacations. An OpenAI engineer described asking the program to look at a Slack conversation about an engineering problem and generate charts, getting back a series of revealing visualizations.
Akshay Nathan, who leads OpenAI’s product engineering team, sums up the value well: There’s a flood of information for the average worker at any of these companies, including me. We’re limited by our ability to process everything that’s available and act on it. That information lives in system-of-record tools. The value of ChatGPT is that you already had access to this, but now you actually have access.
In practice, when the system works, it’s impressive. It was possible to ask ChatGPT Work to extract a poorly formatted preschool calendar from an email and drop it into Google Calendar, saving a ton of repetitive typing. It also built auto-updating financial dashboards for publicly traded companies, created searchable databases, and started sending weekly summaries about new AI research.
It’s not all sunshine: the limits and the learning curve
Despite all that, asking the model something is intuitive, but giving it what it needs to act isn’t quite as simple. Setting up permissions for an agent to access a cloud drive, for example, can be confusing and circular. Many important settings are only available in the web version, which forces users to work in two places at the same time. And some limitations are baffling: connect Google Calendar and the agent creates events, but not new calendars.
Joe Gershenson, who leads OpenAI’s harness team, admits that effort levels still aren’t intuitive for new users. There are things we can do better to help them get to the right level of reasoning, he said, adding a cryptic stay tuned.
There’s an even bigger challenge when it comes to breaking into everyday office work: most workflows aren’t as measurable as code. Software works or it doesn’t. But a good presentation, a business strategy, or a sales pitch aren’t so easy to evaluate. One of the unique challenges of a product like this is precisely that it can do anything, Ambrosino sums up. OpenAI says it uses the GDPval benchmark, covering 44 occupations and hundreds of knowledge-work tests, supplemented by user feedback. ⚙️
The rivalry that shaped OpenAI’s design
Behind all this racing lies a rivalry that no one at OpenAI talks about casually: the competition with Claude, from Anthropic. When asked what sets Codex and ChatGPT Work apart from competitors, the engineers were evasive. The honest answer is that I really don’t look at the harnesses they build, said Gershenson, even pulling out a Mad Men meme.
That’s a little hard to fully believe, considering the similarities between the interfaces and the fact that the first thing ChatGPT Work asks you to do is import your Claude Cowork data. The topic is sensitive because Anthropic defined the AI coding market and kicked off a revolution in how engineers work. It’s even more frustrating because OpenAI had the idea first but didn’t execute it well early on.
When they built Codex as a web app, the engineers were, in Ambrosino’s words, a little too AGI-pilled, betting the model would be smart enough to handle everything on its own. Claude Code, created shortly after, was designed around a back-and-forth conversation with the user, offering options and checking in at every step. Anthropic’s approach proved more effective. Later, OpenAI followed the same path, adding more interaction, and today Codex has taken a slight lead in downloads, with enterprise usage data suggesting the company is catching up to its rival.
So what makes a good harness?
This discussion goes back to the famous bitter lesson learned by AI researchers: a better general model matters more than domain-specific expertise. For the true believers, the harness is a temporary crutch, not the competitive moat. You can get good results in the short term by adding a bunch of extras, but the next model drops in a few months and makes it all obsolete, warns Gershenson.
Not everyone agrees. Ethan Mollick, a professor at the Wharton School who studies AI tools in the workplace, still sees Claude as more user-friendly, writing that ChatGPT tends to want to do magic and solve everything for you, while Claude shows comparisons and asks for feedback repeatedly. Sottiaux disagrees, arguing that the conversational nature of the app is better than learning to use an application. We clearly see that the world seems ready. That’s why we’ve had incredible adoption, he says.
Trust, privacy, and the cost behind the tokens
No conversation about AI agents in real-world environments is complete without addressing trust, privacy, and cost. To work well within a corporate workflow, the agent needs access to sensitive data: confidential emails, internal conversations, strategic documents. That raises serious questions about where that data lives, who accesses it, and how OpenAI ensures it won’t be used to train models without explicit consent.
Then there’s the economic question. In four days of casual use on a 20-dollar-a-month subscription, it was possible to burn through more than 80 million tokens — the equivalent of about 65 dollars, according to the model’s own analysis. That’s a subsidy of more than three times the subscription price for just four days of tinkering. We work every day to push the frontier of efficiency, said Sottiaux, pointing to a recent 80% price cut on one of the company’s models. If you wake up six months from now, you should be able to do the same tasks while spending less.
Mario Zechner, creator of the open-source harness Pi, sums up the challenge of expanding AI agents beyond programming well: Everything has the shape of a coding agent, because training data only exists for coding agent tasks. If I’m a manager and I make a decision today whose outcome only shows up months later, you can’t capture that in a simple back-and-forth log between user and agent. In his view, the big labs push their harnesses as a way to lock in users and avoid becoming mere model providers.
In the end, the promise remains what Nathan calls the magic box. I still think there’s too much complexity, but I’m very optimistic that we can solve it, with the model and in a truly AI-native way, he says. The technology already delivers real value, the business model is structured, and the product is being refined through real-world use. What’s still under construction is the layer of institutional trust that will allow companies of all sizes to adopt these smart workflows safely. And that part, more than any technical breakthrough, will determine whether everyone really ends up using the agents that OpenAI is building. 🛡️
