To what extent are you prepared to grant an LLM authority over your digital existence?
Maximizing the utility of an AI model necessitates relinquishing a significant degree of control. For those who are highly controlling or hesitant about AI, this prospect can seem daunting. However, for Andrew Ambrosino, the lead engineer for OpenAI’s desktop application, it is the sole method to explore future possibilities. Consequently, his app now possesses comprehensive access to and control over his email, Slack communications, phone, and various applications like Notion and Figma, among others.
Ambrosino acknowledged to TechCrunch, “If I instruct it to draft a document, there’s a chance it might draw information from a private direct message on that topic, inadvertently sharing confidential details. Yes, that is a possibility. I will accept that for my work. I am willing to bear occasional personal costs if necessary, though I haven’t yet faced such an issue.”
Ambrosino is involved with OpenAI’s most ambitious project, ChatGPT Work, which was launched last month and is available at the company’s most affordable subscription tier of $20 per month. The aim of this product is to enable professionals in white-collar roles to deploy AI agents, integrating large language models with the digital workflows utilized by accountants, investors, doctors, and anyone else whose daily activities are predominantly computer-based.
OpenAI’s promotional material concisely articulates this objective: to foster a world where “intelligence extends beyond merely answering questions to empowering everyone to actualize their grandest visions.”
While this transformation is already underway for software developers, its adoption in other departments has been gradual. ChatGPT Work is an adapted version of OpenAI’s Codex coding tool, designed to offer non-engineers a similar functionality that software engineers already experience with AI agents: an AI solution that not only answers queries but also independently executes multi-step projects.
Thibault Sottiaux, who oversees OpenAI’s core product development, including Work, explained to TechCrunch, “With this new capability, ChatGPT can autonomously perform entire, highly complex tasks in a manner that is both satisfying and secure. This aligns perfectly with OpenAI’s fundamental mission: to ensure everyone benefits.”
From a business perspective, this is highly significant. Agents that operate for extended periods consume more tokens, making them more profitable for OpenAI per user. Expanding into new professional fields is vital, not just for OpenAI but for the entire industry. While coding has proven to be a profitable area for AI laboratories, it represents only a small fraction of the professional work that AI tools need to facilitate to justify the substantial investments in training and computational resources. While labs have focused on software engineers, specialized competitors like Harvey (for legal services) and Clay (for sales) have been pursuing these customers with a model-agnostic strategy, meaning they integrate whichever AI performs best at a given time.
Industry analysts identify this as a primary hurdle for OpenAI and its rivals. Christian Catalini noted in a16z’s “It’s time to build” blog, “If these laboratories cannot swiftly secure the essential complementary resources required to scale AI in the market, value will instead accumulate elsewhere.”
Enabling AI applications for individuals who are not software engineers demands more comprehensive guidance. Ambrosino recounted that OpenAI’s non-engineering personnel, such as the communications and finance departments, began using Codex when it was “actively unsuited for them — prompting them about code and displaying, ‘oh, you have an empty diff for this thing,'” referring to a technical report for software changes. He added, “Therefore, we began making it more broadly applicable between February and the present.”
A study supported by OpenAI revealed that in June, 98% of OpenAI employees were utilizing Codex, yet only 17% of organizational subscribers and less than 1% of individual subscribers were engaging with the agentic coding tool. This significant disparity between nearly universal internal adoption and negligible external uptake represents both a challenge and an opportunity for the company.
Sottiaux commented, “The greater the value and utility we provide to users, the more willing they will be to pay for a portion of that utility, a principle we have always observed with ChatGPT. You find yourself thinking, ‘Of course, I want to pay $20 a month for this,’ because the value you receive is far greater.”
How to make AI intuitive
To grasp this disparity, it is helpful to understand the foundational work OpenAI’s engineers are undertaking. Every large language model (LLM) necessitates what engineers refer to as a “harness”—the software enveloping the model that dictates the information it processes, the tools it can employ, and how it presents its responses.
For a model to become an agent and perform actions, its harness furnishes it with the necessary tools and instructions for executing long-term tasks. For developers, a command-line interface (CLI) that enabled LLMs to code was sufficient to revolutionize software development and deployment. However, most individuals do not use CLIs; there’s a compelling reason why Windows superseded DOS.
Ambrosino informed TechCrunch that an agentic product extending beyond software engineering “will have to engage with the complex realities of your life, your tools, and websites that were created in 1995 and never updated.” He elaborated that the user experiences his team is developing are crucial for broadening access to beneficial AI.
Consider applications such as Claude Code and Codex: these tools initiated “vibe coding” by abstracting the actual software writing process, allowing users to simply articulate their desired program to the model. Now, OpenAI seeks to make the functionalities found in tools like OpenClaw, which coders use to leverage LLMs, as straightforward as issuing a prompt.
Ambrosino stated, “Without these products providing an interface to the model, experts would still achieve the same outcomes, but widespread adoption by a billion people would not occur.” This inherent tension between the requirements of power users and the demands of mainstream adoption manifests in internal discussions at OpenAI, where some employees contend that a dedicated button is superfluous if users can directly instruct the model.
Ambrosino responded, “We oppose that viewpoint because it is still very early. Discoverability is crucial at this stage, and eventually, the button will no longer be necessary.”
While Work includes additional buttons for selecting projects and plug-ins, it strives for the same intuitive, “magic box” interface characteristic of other OpenAI products. Ambrosino draws a comparison to skeuomorphism, the diminishing design practice of making digital tools visually resemble their physical predecessors, such as a calculator app designed to look like a pocket calculator. He commented, “That wasn’t merely an awkward design choice; it actually facilitated user engagement and eased the transition.”
OpenAI did not disclose the user figures for Work versus Codex, but the combined application is currently utilized by only 20 million individuals, in stark contrast to the over one billion users the company claims are engaging with ChatGPT online.
Giving ChatGPT a license to skill
Presently, OpenAI positions this tool as optimally suited for routine, data-intensive coordination tasks. For instance, its employees are creating weekly metrics reports and converting spreadsheets into planning tools.
I have spoken with venture capitalists who use agents to compile relevant communications and analyses about companies into investment memos, and operations teams who develop custom dashboards and data visualizations. Sam Altman employs it to organize his vacations. An OpenAI engineer described instructing the program to analyze a Slack conversation concerning an engineering issue and “generate some charts,” subsequently receiving a series of insightful plots.
Akshay Nathan, who leads OpenAI’s product engineering team, remarked, “The average worker or employee in any of these companies, including myself, faces an overwhelming volume of information. We are actually quite constrained by our capacity to process all available data and then act upon it. This information resides in various system records tools, such as Salesforce. The true value of ChatGPT is that while you already have access to this information, now you genuinely gain access to it.”
This, then, has the potential to become the digital personal assistant envisioned by AI proponents. Similar to Claude Cowork or the Perplexity AI browsing agent, ChatGPT Work connects agents to your existing digital workspace—including email, web browser, and a multitude of SaaS platforms—and leverages that context to assist you.
When the system operates effectively, it can be quite remarkable: I tasked ChatGPT Work with extracting my son’s irregularly formatted preschool calendar from my email and integrating it into my Google Calendar, which it successfully accomplished, saving me considerable manual data entry. Hopefully, this will prevent me from forgetting school potlucks or neglecting to arrange vacation childcare.
I did not grant OpenAI access to my inbox, interview sources, or story drafts (rest assured, AI skeptics) and would not allow it access to my bank account, although I believe its utility would increase with such trust. I assigned it the task of performing financial analysis on publicly traded companies I cover, and it provided an automatically updating dashboard of metrics; it also created a queryable database of space launches, a task I previously had to achieve by writing Python scripts. Additionally, it sends me a weekly email summarizing new AI research from academic repositories. I intend to continue experimenting with it.
While prompting the model is intuitive, providing it with the necessary information to perform actions is less straightforward. Configuring permissions for agents to access, for example, a cloud drive proved confusing and circuitous—I repeatedly attempted to grant only “read” access and encountered error messages. The model itself offered little assistance, but eventually, on the mobile app, a pop-up dialogue informed me that only full access would enable its functionality.
Many crucial settings are exclusively accessible via the web application, leading me to frequently operate in both environments concurrently. At times, ChatGPT Work’s limitations are perplexing—linking it to your Google Calendar allows event creation but not the generation of new calendars. Furthermore, it is advisable not to attempt any task unless the effort level is set high, otherwise, you will find yourself with the least effective intern imaginable.
This is a common piece of advice from early AI adopters, who worry that frustrated newcomers will abandon the tools. Joe Gershenson, the engineering lead for OpenAI’s harness, conceded that the effort settings are not yet intuitive for new users, stating, “there are areas where we can improve to help them achieve the appropriate level of reasoning…” He concluded by saying, “Keep an eye on this space.”
OpenAI encounters another significant challenge in penetrating mainstream white-collar professions: most workflows are not as quantifiable or assessable as code. Software either functions or it does not, and even that distinction simplifies the nuances of what constitutes good or bad code. Evaluating or tracking the quality of a presentation, business strategy, or sales pitch is considerably more complex.
Ambrosino remarked, “One of the distinctive challenges with a product of this nature is simply its boundless capability.” When I inquired about the specific problems the team designs around and the workflows it targets, the engineers I spoke with hesitated, stating that was a matter for OpenAI’s research division.
OpenAI later provided a response, informing TechCrunch that it employs its GDPval benchmark, derived from 44 occupations and hundreds of knowledge work tests, complemented by user feedback. A less formal answer is that the insights come from OpenAI employees themselves. Ambrosino questioned, “We must consistently analyze… are we engaging in workflows that everyone else will adopt, or are we an anomaly?”
The initial users of the application will generate valuable usage data, much like the success of coding tools relies on similar data collection, provided they do not opt out of making it available for training (which I did).
The rivalry that drove OpenAI’s product design
Despite the considerable focus on the model interface, OpenAI’s engineers were hesitant to answer a rather straightforward question: What distinguishes Codex and ChatGPT Work from Claude Cowork, or other competing agentic harnesses designed for a broad user base?
Gershenson’s typical reply was, “You’re going to find that answer quite underwhelming, and I apologize, but the honest truth is that I genuinely don’t pay attention to the harnesses they’re developing. The ‘I feel bad for you, I don’t think about you at all’ meme from Mad Men comes to mind here.”
Frankly, I find their claims implausible, given the striking similarities in the products’ user interfaces, the essential need for competitive intelligence in any business, and the fact that upon initiating ChatGPT Work, its first request was to import my Claude Cowork data.
It is understandable that Claude Code might be a sensitive topic within OpenAI. Their corporate competitor established the market for AI coding and sparked a revolution in how software engineers conduct their work. This is particularly frustrating because OpenAI initially conceived the idea but failed to implement it effectively.
When OpenAI initially created Codex as a web application, the engineers were perhaps a bit overly ambitious—or, as Ambrosino describes it, “a bit more AGI-focused.” Essentially, they relied on the model being intelligent enough to autonomously handle tasks with minimal user input.
Developed shortly thereafter, Claude Code was designed around an interactive, conversational exchange with the user. If presented with a problem, it would assess the possibilities and offer several options for resolution. Upon selection, it would progress slightly further, then re-engage for continuous updates, thereby minimizing opportunities for errors by the model and its harness.
Anthropic’s methodology proved more efficacious, even if it necessitated greater user engagement. Ambrosino now concedes, “[Our] product was somewhat ahead of the capabilities of the model and harness at that particular time.”
OpenAI eventually adopted a similar strategy by incorporating more opportunities for users to interact with the model. This evolution resulted in the Codex we recognize today, complete with desktop and mobile applications. Using download statistics as an indicator of interest, Claude Code enjoyed greater demand until April of this year, but now Codex has achieved a slight lead; surveys of enterprise usage also suggest OpenAI is closing the gap.
Part of this advantage stems from achieving the correct product-market fit, and another part is due to concerns regarding Anthropic’s model safety restrictions and computational resource shortages. OpenAI’s progression towards a more human-centric harness continues with ChatGPT Work, yet the engineers I interviewed maintained that the primary distinguishing factor is the robustness of OpenAI’s newest powerful and cost-effective models.
Ambrosino stated, “The frustrating truth is that often, it comes down to the model, and one thing we have strived to do exceptionally well with this application is to fully harness the model’s capabilities.”
What makes a good harness, anyway?
That explanation refers back to the “bitter lesson” learned by AI researchers: a superior general model is more critical than specialized domain expertise. For staunch believers, the harness is merely a temporary support, not a lasting competitive advantage.
Gershenson told TechCrunch, “While you could achieve strong short-term results by incorporating numerous additional features—such as ‘if-then’ statements and tools—realistically, the next model will emerge in a few months and render those obsolete.” His team prioritizes the most straightforward methods to expose the model to the necessary tools and context, and nothing more.
Gershenson elaborated, “The objective of effective harness engineering is to… be more precise about the specific information the model truly requires to resolve your problem, because the models are continuously improving at accomplishing that if you simply allow them to operate independently.”
Nevertheless, there remains an unresolved question regarding whether most individuals or models are prepared for such an approach. Ethan Mollick, a professor at the Wharton School of Business who studies AI tools in professional settings, continues to view Claude as more user-friendly, observing that “ChatGPT tends to aim for seamless automation, performing tasks for you directly, whereas Claude engages in comparisons, presents options, consistently requests input and feedback, and conducts A/B tests.”
Sottiaux, and potentially OpenAI collectively, hold a differing view, contending that the conversational nature of the application is preferable to having to learn how to operate a new program. Regarding the app, he asserts, “We are confident that the world appears ready. This is precisely why we have experienced remarkable adoption.”
Despite this, it remains unclear whether a model-specific harness is the most effective strategy for maximizing a model’s potential. Comparative analyses conducted by companies such as Composio and Databricks demonstrate that various combinations of harnesses and models yield differing performance on coding benchmarks. Databricks discovered that Pi, an open-source harness released by the software firm Earendil, surpassed Codex in performance while utilizing the identical GPT 5.5 model. Pi has been instrumental in developing projects like OpenClaw and Cloudflare OS.
Mario Zechner, the creator of Pi, contends that his deliberately minimalist harness proves that an AGI-centric approach can be effective, at least for software engineers and coding tasks. He states that its perceived lack of explicit features is compensated by its capacity to self-modify and construct its own interfaces. He empathizes with the difficulties OpenAI’s engineers face in broadening their user base beyond engineers.
He explained to TechCrunch, “Everything is structured like a coding agent… the reason is that they exclusively possess training data for coding agent tasks. For instance, if I, in management, make a decision today, and its outcome manifests months later, that cannot be captured in a simple back-and-forth trace between a user and an agent. Consequently, all such tasks and anything that isn’t digitized remains inaccessible for a model to learn from.”
Similar to other open-source providers, he perceives the major labs’ efforts to promote their proprietary harnesses as a strategy to lock in users; he argues, “They must control the entire technology stack; otherwise, they risk becoming merely a model provider and then have to compete with Chinese models.”
He and other engineers interviewed by TechCrunch felt that the visibility into token consumption and agent behavior within the frontier labs’ harnesses is too limited. While this may be less significant for non-technical workers, the widespread adoption OpenAI anticipates will eventually necessitate more rigorous discussions about costs, akin to the situation with coding tools.
For example, during four days of casual experimentation with a $20-per-month subscription, I consumed over 80 million tokens, which the model’s analysis indicated cost $65 (the app lacks a dashboard). This represents a subsidy of more than three times the subscription price for just four days of light usage.
Sottiaux noted, referencing a recent 80% price reduction for users of OpenAI’s Luna model, “We are diligently working daily to advance the frontiers of efficiency. If you were to assess the situation six months from now, you should be capable of performing all the identical tasks with reduced expenditure.”
Another pertinent question is whether these applications create a vendor lock-in effect for customers, either through data retention practices or the significant difficulty involved in configuring access for all plug-ins and their associated permissions.
During my visit in July to OpenAI’s headquarters, characterized by its wood-paneled walls and abundant plant life, the atmosphere, though calm, felt subtly charged; these individuals clearly have extensive work to complete. The engineers I spoke with continuously monitored their laptops while conversing and moved briskly from one meeting room to another.
Nathan, who heads the product engineering team, affirmed that the focus remains on “the promise of the magic box,” but added, “I still believe there’s excessive complexity. I am highly optimistic that we can resolve this, utilizing the model and through a genuinely AI-native approach.”