A guest post from Thomas Prommer, who’s held senior technology leadership roles at Adidas and Huge Inc.
I work from a lot of different places throughout the day. Coffee shops, airport lounges, hotel desks during the week, pair programming sessions next to my developers when I am in the office. People see my screens. And nearly every day, at least one of them leans over and asks some version of the same question: what is all that running text on the black screens, and where is your Outlook, your Slack, your Excel, your to-do manager, your browser?
The honest answer is that for most of my actual working day, I have moved past those applications. The black terminal is where I read tickets, draft documents, query databases, write briefs and reply to email. It does that through agents that talk to a growing stack of command-line interfaces, Model Context Protocol servers, APIs and connected internal data sources. The applications I used to live in are still installed somewhere. I touch them less every week.
This is, in my view, the interim shape of a new operating system: AI-driven from the ground up, where work happens through agents and connected data rather than through clicking around an application UI. It is the leading indicator of a shift that belongs on every technology leader’s whiteboard. The operating system, as we have known it for forty years, is being quietly replaced. Not the kernel. The layer above the kernel that we actually live in, the menus, windows, inboxes and ribbons, is becoming optional.
What is actually being displaced
The visible OS is the application layer. Outlook, Word, Excel, Chrome, Slack, Figma. For decades this is where work has happened. A human points and clicks, the application responds, the human composes the next click. It is a beautiful system and it has been remarkably durable.
Andrej Karpathy describes the change as a move from “typing lines of code” to “delegating larger macro actions” such as implementing a feature, refactoring a subsystem or running tests. He notes that around December 2025 he began trusting agents with progressively more of the work. The same pattern shows up in production now. Simon Willison has been documenting through 2026 how terminal-native agents quietly became part of how senior developers actually do their jobs. The act of using software now reorganises itself around an agent that uses software on your behalf.
Software developers are leading the pack here, which is unsurprising: their work was already text-shaped and tool-rich. My own view is that this will spread far more broadly into non-technical roles. I am starting to see it happen across functions, not just engineering, because the productivity delta against an application-first workflow is no longer subtle.
If the agent is doing the clicking, then menus, ribbons and dashboards exist to serve a user who is no longer there. The application UI becomes scaffolding for a worker that does not need it. That scaffolding has weight. It costs you money. It costs your vendors innovation cycles. And it is now in the way.
Graphical interfaces will not disappear. They survive wherever a human still needs to see something, sign off on it, or work visually, and that covers a lot: design tools, financial models, anything where a wrong decision is expensive. What changes is the default. Two years ago the question was which application to open. Now it is which agent to ask, and whether it has the access it needs.
The new substrate is command-line tools, agents, APIs and MCPs
In December 2025 Anthropic, Block and OpenAI donated their core agent projects to a new Agentic AI Foundation under the Linux Foundation, with support from Google, Microsoft, AWS, Cloudflare and Bloomberg. The Model Context Protocol was the headline contribution. In its first year MCP has grown past 10,000 active servers and 97 million monthly SDK downloads, with first-class client support across ChatGPT, Claude, Cursor, Gemini, Microsoft Copilot and VS Code. This is not a niche developer tool. It is the way agents now reach data, tools and other agents.
Three layers are settling. MCP gives agents tool access. Agent-to-agent protocols handle coordination. Streamable HTTP carries the traffic. Anything that wants to be useful in this stack has to be reachable through it. Anything that is not risks becoming invisible to the agent layer: a well-designed app whose buttons no one will ever press, because no human is sitting in front of it.
This is why the 2026 office-suite battle is so revealing. Microsoft has rebuilt Copilot inside Word, Excel and PowerPoint. Google has reframed Workspace around an “office intern” agent. On 7 May Anthropic shipped Claude as add-ins for Excel, Word, PowerPoint and Outlook, with a single agent that preserves context as it moves between the four applications. Gartner places $58 billion at stake in this productivity-suite reshuffle, the first credible challenge to the category in thirty-five years. None of this is a closed case. Productivity-suite incumbents have survived prior waves, and it is fair to ask whether the agent layer is genuinely structural or another set of features bolted onto familiar shells. The honest read is that we are early, and the leading indicators are pointing one way.
The governance layer comes with it
A new operating layer creates new failure modes, and senior leaders should plan for them now. Agents propagate errors at machine scale and with less intuition than a human about consequences. Permissions need to shift from “what can this user see” to scoped, time-bounded, auditable agent identities. Every MCP server in production should have an authorisation model and an activity log a regulator could read. Watch for agent sprawl too. If every team spins up its own MCP servers against its own data, you have rebuilt shadow IT in a more powerful shape. Get identity and observability under one roof now, while the surface is still small. Cost discipline belongs in the same conversation. Per-inference billing is closer in shape to a utility bill than a software licence, and organisations that move first without a FinOps mindset for AI will be paying for the lesson for several quarters.
Expect a vendor reshuffle
A new OS will need a new marketplace, and several already exist: Claude Skills, the GPT Store, MCP hubs such as mcp.so and Smithery, Replit’s agent market and Cloudflare’s AI Marketplace, alongside earlier infrastructure like Hugging Face Spaces and the LangChain Hub. Pricing models are diverging (revenue share, free distribution, direct sale, per-inference billing) much as they did in the early app-store era. None has won. One or two will.
It would be unusual if the incumbents we have lived with since the late 1990s emerged from this transition with the same market positions. New gatekeepers tend to appear when the surface that users live on changes. That surface is changing now.
What to do this quarter
Three moves, in order of how reversible they are.
The first is to treat agent governance as a real workstream this quarter, not a footnote on someone else’s roadmap. That means scoped, time-bounded agent identities; an audit log for every MCP server in production that a regulator could read end to end; and a FinOps view of inference cost from day one. Set the policy now, while the deployment surface is still small.
The second is to make your internal systems addressable by agents. Audit which of your tools expose real APIs and MCP servers. The ones that do not should get a small project to expose them. Treat this the way you treated mobile responsiveness in 2012: an unglamorous foundation that decides who gets the next decade.
The third is to hold any large multi-year licence with one eye on the door. This is not a call to rip and replace. It is a call to negotiate shorter terms, retain optionality, and avoid being the organisation that signed a seven-year deal with the wrong vendor in the year the OS quietly changed.
The black terminal window on a senior engineer’s screen is not an aesthetic choice. It is the early shape of what most knowledge work will look like, at least in the parts of the business where the work is fundamentally information rather than judgement. The leaders who see that, and rebuild their systems, governance and supplier choices to match, will be in a very different position by this time next year than the ones who do not.
Thomas Prommer
Thomas has held executive tech leadership roles as Global SVP of Engineering at Adidas, CIO at Sweetgreen and President of Technology at Huge Inc., where he led digital transformation for Apple, McDonald’s, Google, Nike, Toyota, Four Seasons and Carnival. He has scaled engineering organisations of more than 1,000 people across multiple countries.
