From AI 'Answering' to 'Executing': How Codex is Remaking the OS of Work
Saying 'AI has become smarter' is not enough to fully transform work. What is striking about this conversation (a session within Cisco's AI Summit) is that Sam Altman is shifting his focus from 'model progress' to 'usage (interfaces) and organizational adoption.' When AI moves from being a 'tool that answers' to a 'colleague that executes,' where do the real constraints lie for companies? I will organize the points based on his remarks.
1. 'Codex is the new ChatGPT moment'—An experience that crossed the threshold
Altman stated that OpenAI's Codex has followed an 'exponential curve' over the past few months, emphasizing that the feel of product development has changed. In fact, a speaker from Cisco mentioned that 'almost all AI Defense code is written by Codex,' suggesting that the main role in development is shifting beyond traditional 'auxiliary code generation.'
Altman described this feeling as 'another ChatGPT moment.' The point is not just the performance of the model, but that 'the UI and the harness (the mechanism for mastery) have caught up.' In other words, the 'last mile' where capability is converted into product value has been bridged.
2. From tool to 'colleague': The meaning of AI operating a PC
This is the core of the conversation. Altman stated that 'combining code with general-purpose computer use (browser operation, document editing, research) makes it far more powerful,' describing a future where AI agents enter 'your environment' to complete tasks. He even shared an anecdote: 'At first, I didn't intend to allow full control, but it was so convenient that I gave in after two hours,' and 'That's why I'm using two laptops now.' It is a common workplace reality where convenience overrides decisions on secure operation.
The Codex app for macOS was released as a product that pushes this 'colleague-ization.' It is described as aiming to be a 'command center' that runs multiple agents in parallel, delegates long-term tasks, and moves forward while you review them.
“You have to think of it not as a 'tool' but as a 'teammate.'”
This statement also directly relates to the KPI design on the adoption side. It is not about increasing the number of people sitting at desks, but because you need a design that grants authority to colleagues (AI).
3. Non-obvious constraints: Security, authority, and 'remaking software'
'The constraints are power and computing resources'—everyone says that. What Altman cited as 'non-obvious' is actually the root of corporate operations.
3-1. A new compromise between data access and security
Altman says that 'security and data access vs. utility' is something 'no one has solved yet,' and that 'a new kind of security/authority paradigm is needed.' While AI becomes valuable enough to use 'a browser with your session,' the quality of risk changes at that very moment.
3-2. From 'UI for humans' to 'design for agents'
What is interesting is that Altman suggested the need to 'remake all software to be equally easy for both humans and AI to use.' The Slack example is symbolic: if an agent reads through threads in a Web UI, unread messages become read, breaking the workflow. He says this requires 'separate accounts for AI' or designs that are 'primarily used by AI but can still be used by humans in the traditional way.'The implication is that the UI is not the main battlefield; rather, authority, auditing, APIs, and state management are.
4. 'Capability overhang' and the gap in adoption speed
What Altman repeats is the gap between 'what is possible' and 'widespread adoption.' While ChatGPT and Codex themselves are growing fast, corporate absorption is slow. The keyword that appears here is 'capability overhang' (a state where capabilities are excessively ahead of the actual workplace).
As a prescription, he emphasizes organizational design over technology.
Resolve the 'clogs' in security/data access early.
Accelerate adoption decision-making with 'AI colleagues' as a premise.
The possibility that 'worrying for a year before adopting' will be too late.
And as a slightly bold prediction, he states that "companies that cannot quickly hire AI colleagues will be at a major disadvantage." This is not mere fear-mongering, but a point about a structural shift where competitive advantage moves from "model performance" to "speed of adoption."
5. Infrastructure and Open Source: Demand "Increases as Costs Decrease"
On the infrastructure side, Altman notes that "models will become cheaper and run on fewer resources," but because "people will use them more as they become cheaper," total demand will continue to grow. He explains that market demand is like "electricity," where the demand curve shifts based on price and quality (intelligence and speed).
He goes even further on the topic of open source. While expressing concern about the U.S. lagging behind in open source, he says that while the frontier (APIs, etc.) is the top priority, demand for "locally running private models" will increase. This is because if AI is to see "one's entire life," the value of it running locally and being controllable by the user will rise.
Conclusion
What this conversation shows is not just a "future where models become smarter," but a future where AI becomes an "execution agent," remaking the corporate OS (permissions, auditing, software design, and implementation decision-making).
A realistic step that readers can take starting today might be to articulate the following two points before jumping into flashy PoCs.
The scope of authority that can be granted to AI (data access, auditing, and division of responsibility)
How to decompose tasks assuming an "AI colleague" (focusing on state management and APIs rather than UI)
The anecdote that "it was so convenient that the policy collapsed in two hours" is not a story of the future, but the very "reality of operations" that has already begun in the field.