OpenAI bets on AI agents for everything, and ChatGPT Work is the test

OpenAI is trying to take AI agents beyond developers. In a TechCrunch report, the company detailed ChatGPT Work, a product launched the month before publication and available on the cheapest plan, at $20 per month. The idea is to give accountants, investors, doctors and other office professionals the same capability engineers already have with coding agents: a tool that not only answers questions but executes multi-step tasks autonomously.

The technical detail relevant to those following the topic: ChatGPT Work is a modified version of Codex, OpenAI's own coding tool. In other words, the company is taking the architecture that worked for engineers and adapting it for everyone.

What the Numbers Reveal

A study backed by OpenAI ("The Shift to Agentic AI: Evidence from Codex") showed the size of the gap between internal and external use. In June:

• 98% of OpenAI employees used Codex;
• only 17% of organizational subscribers;
• less than 1% of individual subscribers.

The combined product (Work + Codex) is used by about 20 million people, compared to more than 1 billion who use ChatGPT with conventional prompts. This gap between near-total adoption in-house and low adoption outside of it is, at the same time, the company's challenge and opportunity.

Commercially, the bet makes sense: agents that run for longer periods consume more tokens, which makes them more profitable per user. And coding, while profitable, is a small subset of the professional work that labs need to automate to justify the billions invested in training and compute.

Why Moving Beyond Code Is Hard

Every LLM needs a harness, the software around the model that decides what it sees, which tools it can use and how it returns answers. For developers, a CLI that let the LLM code was already enough to change the way software is built. But, as Andrew Ambrosino, chief engineer of OpenAI's desktop app, put it, most people don't use a CLI, and there's a reason Windows replaced DOS.

An agent that goes beyond software engineering, Ambrosino said, will "play around in the messy world of your life, your tools, and sites built in 1995 that were never updated." He himself gave the app access to his own inbox, Slack, phone, Notion and Figma, accepting the risk: "I'm going to take the personal hit every now and then if I need to. And I haven't needed to."

There's also an evaluation problem. Software works or it doesn't, but a good presentation, business strategy or sales pitch isn't as easy to measure. To calibrate this, OpenAI said it uses the GDPVal benchmark, built from 44 occupations and hundreds of knowledge-work tests, complemented by user feedback.

How It Works in Practice

The TechCrunch reporter described concrete use cases. He asked ChatGPT Work to extract his son's preschool calendar, poorly formatted and stuck in an email, and put it into Google Calendar, a task that worked. He also generated an auto-updating metrics dashboard for publicly traded companies and built a queryable database of space launches, something that previously required writing Python scripts.

But setup still stumbles. Granting "read" permission to a cloud drive resulted in repeated errors, and only a dialog box in the mobile app explained that only full access would work. Many settings only exist in the web version, forcing users to switch between platforms. Linked to Google Calendar, the agent creates events but doesn't create new calendars. And the recurring advice from early users: without the high effort level, "you have the worst intern you've ever worked with."

The Rivalry with Claude Behind the Design

The report highlights that the product's design was shaped by competition with Anthropic. OpenAI had the idea first, launching Codex as a web app, but bet heavily on model autonomy, becoming too "AGI-pilled," in Ambrosino's words. Claude Code, built shortly after, revolved around back-and-forth conversation: it presented three or four options, waited for the user's choice, moved forward a bit and checked in again, leaving less room for error.

Anthropic's approach proved more effective. Using downloads as a proxy for interest, Claude Code led until April, but Codex took a slight lead afterward, also helped by complaints about safety restrictions and lack of compute on Anthropic's side. Ethan Mollick, a Wharton professor who studies AI at work, still sees Claude as friendlier: "ChatGPT tends to want to do magic and do everything for you, while Claude compares and shows, repeatedly asking for input."

The Debate over the Harness and the "Bitter Lesson"

There's a technical discussion here that matters to anyone building products with LLMs. OpenAI's engineers insist that the real differentiator isn't the harness, but the model's strength and cost-effectiveness, invoking the "bitter lesson": a better general model is worth more than domain-specific expertise. For Joe Gershenson, OpenAI's harness lead, filling the software with "ifs, thens and tools" is a waste of time because "the next model comes out in a few months and makes it obsolete."

But the data complicates the thesis. Comparisons by Composio and Databricks show that different combinations of harness and model deliver different performance on code benchmarks. Databricks found that Pi, an open source harness published by Earendi, outperformed Codex using the same GPT 5.5 model. Mario Zechner, creator of Pi, points to the structural limit: "everything is shaped like a coding agent, because they only have training data from coding agent tasks." Management decisions, whose outcome only appears months later, don't fit into a simple user-agent conversation trace.

Zechner also sees the big labs' insistence on pushing their own harnesses as a lock-in strategy: "they need to own the entire stack, otherwise they become just a model provider and then have to compete with Chinese models."

What This Means for the Brazilian Developer

For those building software in Brazil, the practical message is twofold. First, the agent architecture that already changed the coding workflow is being generalized to other areas of the company, meaning product roadmaps will start to consider agentic automation in sales, operations, finance and business areas, not just engineering. Second, the harness-versus-model debate has direct consequences for stack choices: vertical competitors like Harvey (legal) and Clay (sales) adopt a model-agnostic approach, plugging in whatever works best at the moment, while OpenAI is betting on coupling its own product and model. The evidence that a minimalist, open source harness can beat the official one using the same model is an argument in favor of architectures less tied to a single vendor.

It remains an open question whether non-technical audiences are ready to hand over the keys to their digital lives to an LLM, and whether OpenAI's bet on a proprietary harness will hold up as models evolve.