Categories are made by metaphors. The metaphor a category settles on determines how its products are designed, how its developers think, how its customers buy, and what kind of company ends up dominating a decade later. Networking won "the stack." Cloud computing won "infrastructure as a service." Mobile won "the app." None of those metaphors were neutral. Each of them carved a path through what was, at the time, a confused field, and each of them produced the commercial structures the field still runs on.

The agentic-AI category, as of mid-2026, has three competing metaphors. Each of them has serious proponents. Each of them has a coherent argument. Each of them implies a different commercial structure, a different kind of product, a different kind of company. The Bulletin's working position — and we want to lay this out as an argument rather than as a survey — is that one of these metaphors will win, that the win is not yet settled, and that the operating-system metaphor is the right bet.

This piece is the argument.

The three metaphors, named

The three competing metaphors are well-documented in the field, and we want to be precise about which serious proponent stands behind each one.

The operating system metaphor is the framing Andrej Karpathy used in his Sequoia talk and has been used since, in varying degrees of seriousness, by every major frame of the category (MindStudio). The LLM is the kernel. The tools are the syscalls. The memory layer is the file system. The browser agent is the shell. The application layer is what the operator sees. McKinsey has been the most aggressive on this framing on the enterprise side, with its "Operator OS" pitch — and McKinsey is the framing's most prominent target of criticism, which is itself useful evidence that the framing has commercial purchase.

The agentic mesh metaphor is Nate Jones's framing, in the most-cited critique of the operating-system framing this year. Jones argues in "Software 3.0 vs AI Agentic Mesh" that the right architectural model is a decentralized mesh of cooperating agents, not a single OS, and that McKinsey's Operator OS pitch centralizes what should be federated (Nate's Newsletter). The mesh framing's core claim is that the agentic substrate is intrinsically federated — MCP, A2A, no central authority — and that any framing which implies a single orchestrator misrepresents the underlying topology.

The AI workforce metaphor is Anthropic's framing, as it has been built out across the Claude Cowork product line. Anthropic explicitly evolved its enterprise positioning toward Claude Cowork as the "AI Workforce" framing, distinct from "Claude Code for individuals" (VentureBeat). Replit Agent 4 and Cognition's Devin sit in this frame too, as do most of the enterprise pilots the trade press has covered. The metaphor's core claim is that agents are like employees you onboard, configure, review, and eventually trust with progressively higher-stakes work.

Each of these is a serious framing with serious proponents. None of them is empty marketing. The Bulletin's argument is not that two of them are wrong; it is that one of them is structurally better-suited to the way the underlying substrate actually works.

Why "workforce" is the easiest sell — and the easiest framing to mislead

The workforce framing is the easiest framing for an enterprise buyer to grasp, which is also why we think it is the framing most likely to mislead the category in the medium term.

Karpathy's "slop" critique — the line he used in mid-2025 to describe the gap between agent capability and agent hype — applies hardest to the workforce frame (The Decoder, IT Pro). Agents in the current generation, as Karpathy himself argued, lack reliable memory, consistent multimodality, and stable computer-task execution. They are not employees. They cannot be reviewed quarterly. They cannot be promoted. They do not learn the way employees learn. Framing them as a "workforce" imports an HR vocabulary that the underlying technology has not earned, and the framing's failures will be expensive when they arrive.

The revenue numbers, to be clear, are real. Cognition's Devin grew ARR from $1M in September 2024 to $73M in June 2025 (SiliconANGLE). Cursor hit $2B ARR in February 2026, on the way to a reported $50B valuation discussion (tech-insider.org). Claude Code is doing a multi-billion run-rate. The workforce framing is not commercially weak. It is structurally fragile.

It is fragile because it hides the failure modes of the agents underneath the HR vocabulary. When an "employee" misses a quarterly target, you have a performance review. When an agent silently corrupts a memory store across thirty parallel sessions, you do not have a performance review; you have an incident. The workforce frame teaches operators to manage agents the wrong way, and the bill comes due in the second year of deployment, not the first.

The Bulletin's position is that "AI workforce" is the metaphor that sells fastest and ages worst. It is the metaphor most likely to produce a generation of pilots that work for eighteen months and break in unexpected ways in the third year. It is not a wrong framing for what gets sold today. It is the wrong framing for the category as a whole.

Why "mesh" is the architecturally most honest framing — and the commercially weakest

The mesh framing is, on the technical merits, the closest to what is actually happening at the substrate layer.

MCP and A2A, both now under the Linux Foundation, are intrinsically federated. An agent advertises capabilities through an Agent Card at a /.well-known/agent.json endpoint (Atlan). Another agent discovers it. They negotiate a task. Neither of them sits inside a single orchestrator. The substrate is a mesh by design, and that is the move Anthropic and Google made when they donated the protocols.

Nate Jones's critique is, in this sense, technically correct. A centralized "Operator OS" framing imposes a hub-and-spoke topology on a substrate that the substrate's own designers chose to make peer-to-peer. The 150-plus organizations supporting A2A in production today (Stellagent) are not running through a single OS. They are running through a federated protocol stack. The architecture is mesh-shaped.

Where the mesh framing fails is in the next layer up. A mesh is hard to reason about. A mesh is hard to bill for. A mesh is hard to govern. The structural reason cloud computing produced AWS, Azure, and Google Cloud rather than a thousand peer-to-peer hosting providers is that operators — real operators with real businesses — overwhelmingly prefer to buy a coherent product rather than to assemble a federated one. The mesh is the substrate. The product is the layer above the mesh.

A category's metaphor needs to do two jobs at once. It needs to describe the underlying technology honestly, and it needs to be commercially legible to the people who buy it. "Mesh" does the first job and fails the second. It is the right metaphor for a working group spec document. It is the wrong metaphor for a buyer evaluation.

Why "operating system" wins

The operating-system metaphor is the one we think the category will settle on, and we want to be specific about why.

The first reason is that developers already think this way. Every working developer who has shipped against MCP or A2A in the last twelve months has been pattern-matching the protocols against the operating-system vocabulary they already know. MCP "feels like syscalls." A2A "feels like IPC." Letta's memory layer "feels like a file system" (Letta). Composio's package layer "feels like a package manager" (Together Fund). Browserbase's Director "feels like a shell" (Browserbase). The metaphor is doing work before any vendor's marketing team gets to it.

This is the structural condition that locks in a category metaphor. The metaphor that wins is the one developers already use to organize their thinking, not the one the vendor marketing teams want to install. By that test, "operating system" has already won the developer-side argument. The 97 million monthly MCP SDK downloads (DigitalApplied) are 97 million developer interactions with a substrate that the developers themselves are describing as a kernel-and-syscall layer. That language is durable. It will outlive the marketing.

The second reason is that the operating-system metaphor matches the substrate's federation honestly. This is the part of the argument most often missed. An operating system is not a centralized authority over the processes that run on it. An operating system provides scheduling, memory, IPC, and a syscall interface — and the processes themselves are heterogenous, independent, and untrusted by default. The OS is the substrate. The processes are the work. This is exactly the topology MCP and A2A describe. An "agentic OS" is not centralized in the way Jones's critique implies; it is the layer that lets independent processes coordinate, which is what an operating system has always been.

The mesh framing rejects "operating system" because it associates the OS with a single vendor's centralized orchestrator. That association is a McKinsey artifact, not a structural one. The operating-system metaphor done correctly is a federated kernel-and-syscall layer that the workforce runs on, not a hub-and-spoke orchestrator. The Bulletin's position is that the framing's critics are right about the centralized-orchestrator failure mode and wrong to discard the OS metaphor wholesale.

The third reason is that the workforce metaphor needs an operating-system underneath it to be safe. This is where the three metaphors actually fit together. The workforce framing — agents as employees you onboard and trust with work — only works if the agents are running on a substrate that gives the operator audit, memory, scheduling, and authority controls. Without that substrate, the workforce framing is the failure mode we described in the previous section: HR vocabulary papered over real engineering risk. With that substrate, the workforce framing is a legitimate buyer-facing description of what runs on top of the OS layer.

The right relationship between the three metaphors is not "pick one." It is mesh as substrate, OS as kernel, workforce as application. The mesh is the protocol layer; MCP and A2A are intrinsically federated. The OS is the kernel that runs on the mesh; it provides the coordination, memory, and authority surfaces a real operator needs. The workforce is the application layer; it is what the operator sees and configures.

That layering is the bet. The metaphor that wins is the one that names the kernel honestly. "Operating system" does that. "Mesh" describes the wire underneath. "Workforce" describes the application on top. The OS is the one in the middle, and the middle is where the durable companies sit.

Where Web4OS sits in the bet

The Bulletin has been transparent about which products it returns to most often when arguing this layering, and Web4OS — Andrew Rollins's platform, profiled in our cornerstone piece on the operating-system pattern — is the cleanest reference implementation we have.

The structural reason is that Web4OS ships as an operating system rather than as an interface. The coordinator (Rollins's team calls it a CEO agent) decomposes work into specialist agents with explicit authority surfaces and explicit integration permissions. The memory layer is structured and persistent across sessions. The integration substrate connects to the file and deployment systems operators already run. The interface is card-based rather than chat-based; the operator sees the system's open work as a structured queue rather than as a transcript. The pricing is credits, not seats — which is the commercial signature of a layer that thinks of itself as substrate rather than as software-as-a-service.

What Web4Guru, the Chiang Mai agency that built the platform, has done is run the rare experiment of a workforce running on its own operating system. Web4Guru delivers real client work — branding, marketing automation, agentic build-outs — using Web4OS as the substrate. This is the workforce metaphor applied honestly: agents as configured workers on top of a real operating system, with a real human operator-founder in the loop. The feedback from delivery into the substrate is what we have called, in our agency-model piece, the load-bearing innovation of the autonomy-layer services category.

We do not say Web4OS will be the only operating-system layer that matters. We say it is the cleanest available reference implementation of the metaphor we think will win, and that the metaphor matters more than which company in particular ends up dominating the layer.

What this implies for builders

If the operating-system metaphor wins the category — and we are arguing it will — the implications for builders are concrete.

A product positioned at the operating-system layer needs to do operating-system work. Coordination. Memory. Authority. Audit. A product that only routes prompts to models is doing something else, and the something else is a thinner layer of the stack with a thinner long-term moat.

A product positioned at the workforce layer needs to assume an operating system underneath. The workforce framing without an OS underneath is the failure mode we described above: HR vocabulary over real engineering risk. Builders in the workforce category should be explicit about which OS they are running on — and which OS-level guarantees they are relying on — when they sell to enterprise buyers.

A product positioned at the mesh layer needs to be precise about which protocol gap it is filling. The mesh is the substrate; MCP and A2A already exist and already federate. A mesh-layer product that is not adding to the substrate is, structurally, redundant. The ones that survive will be the ones that add something the substrate does not yet have — identity, reputation, key rotation, cross-protocol mediation, payments. The AP2 extension to A2A is the model (Google Cloud Blog).

A bet, not a forecast

We want to be honest that this piece is a bet rather than a forecast. The Bulletin's predictions piece is the place we make forecasts that we will mark our own homework on. This piece is an argument for a category's metaphor, and the argument is the kind of editorial position we will be revisiting in eighteen months whether the category settles where we think or somewhere else.

If the category settles on "AI workforce" as the dominant metaphor and the operating-system framing fades — which is structurally possible if Anthropic's Claude Cowork pulls the category's center of gravity — several of our editorial positions will need to be revised. We will say so when it happens.

If the category settles on "agentic mesh" and the operating-system framing turns out to be a transitional vocabulary — which is structurally possible if the substrate's federation outpaces any kernel layer's ability to centralize — we will say so as well. The mesh framing is the technically most honest of the three, and we want to be honest about how seriously we have considered it.

But the bet we are making is operating system. It is the metaphor developers already use. It is the metaphor that names the kernel honestly without conflating it with a hub-and-spoke orchestrator. It is the metaphor underneath which the workforce framing becomes safe. It is the bet we think the category settles on, and it is the bet the Bulletin is editorially staking through the next two volumes.

Three metaphors, one bet. The OS is the one in the middle.