We need an Agent OS

The opinions expressed in this article are my own and do not necessarily reflect those of my clients or employer.

Every wave of technology arrives with a grand promise. For AI agents, that promise is simple: they will take work off our hands. These intelligent systems—designed to reason, plan, and act autonomously across applications—are supposed to finally bridge the gap between information and action. The demo videos are mesmerizing. We see agents booking travel, generating reports, and handling entire workflows end-to-end. But in the real world, agents are far better at helping us do the thing than doing the thing themselves.

This is not simply a question of model capability or technical maturity. It is rooted in human psychology. Throughout history, humans have embraced tools that amplify our abilities, but we hesitate to adopt tools that remove our presence entirely. The spreadsheet did not replace the analyst; it made the analyst faster. Autopilot did not eliminate the pilot; it made flights safer. Even Google Search, the ultimate productivity amplifier, leaves the final click and decision to us. AI agents are colliding with the same truth: we are comfortable with co-pilots, but we are not yet ready for captains.

The Bull and the Bear case for AI Agents

The AI community today is split between those who believe agents are on the verge of revolutionizing work and those who see them as fragile experiments.

The optimists, including researchers like Andrej Karpathy and many early enterprise adopters, imagine a near future where agents reliably orchestrate multi-step tasks, integrating across our apps and services to unlock massive productivity gains.

The skeptics, from voices like Gary Marcus to enterprise CIOs burned by early pilots, argue that most agents are brittle, overconfident, and ultimately stuck in “demo-ware” mode—fun to watch, hard to trust.

Both perspectives are correct. The bulls see where technology could go, while the bears are grounded in how organizations actually adopt and govern new capabilities. The tension between promise and practice is not about raw intelligence alone. It is about trust, reliability, and the invisible friction of human systems that are not yet designed for fully autonomous software.

What This Means for Leaders and Builders

For a business executive, the lesson is clear: staking a transformation on fragile autonomy is reckless. The path forward is to pilot agents as co-pilots embedded into existing workflows, where they can create value without assuming total control. Early wins will come not from tearing down processes but from enhancing the teams you already have.

For a technology leader, the challenge is architectural. Building a future-proof approach means resisting the temptation to anchor on a single agent vendor or framework. Interoperability and observability will matter far more than flashy demos. Emerging standards for tool integration and context sharing will shape how agents scale safely, and the organizations that anticipate these shifts will be ready when the hype cycle swings back to reality.

And for the builders—the developers crafting agents today—the priority is survival in a fragmented ecosystem. Agents that are resilient, auditable, and easy to integrate will outlast magical demos that break on first contact with messy enterprise data. Human-in-the-loop design is not a compromise; it is a feature that earns trust and keeps your agent relevant.

The Realization: Agents Need a Home

As I worked through these scenarios, one realization became inescapable: agents are not failing because they cannot think. They are failing because they have nowhere to live. They lack persistent context across tools and devices. They lack a reliable framework for orchestrating actions, recovering from errors, and escalating when things go wrong. And most critically, they lack the trust scaffolding—permissions, security, and auditability—that enterprises demand before they let any system truly act on their behalf.

Right now, each agent is an island. They run in isolated apps or experimental sandboxes, disconnected from the unified memory and control that real work requires. This is why the leap from co-pilot to captain feels so distant.

History offers a clue about how this might resolve. Every major paradigm in computing eventually required an interface and orchestration layer: PCs had Windows and macOS, smartphones had iOS and Android, and cloud had AWS and Azure. Each of these platforms provided not just functionality but a home—a trusted environment where applications could operate safely and coherently. AI agents will need the same thing.

Call it, perhaps, an Operating System for Agents. We may not name it that in the market, but functionally, that is what must exist. The companies best positioned to build it are not the niche agent startups or the AI model providers, but the giants that already own our daily surfaces and our context: Microsoft, Apple, and perhaps Google. They control the calendars, files, messages, and permissions that agents will need to function meaningfully. Whoever creates the layer where agents can live, learn, and act—while keeping humans in control—will define the next era of AI.

The Quiet Future Ahead

If this plays out as history suggests, the rise of agents will not feel like an overnight revolution. It will feel like a quiet infiltration. Agents will first live as co-pilots within our operating systems, helping us with tasks at the edges of our attention. Over time, they will orchestrate more of our digital lives, bridging apps and services in the background. And finally, the interface itself will become the true home of agents—a layer we barely notice, because humans will still remain in the loop, just as we prefer.

The agent future, then, is not about replacing us. It is about designing the world where they can live alongside us. Only when that world exists will the real revolution begin.