Playbook

AI agents are starting to be built like video games. Two announcements, a week apart, made it hard to miss. On October 1, Anthropic let Claude Code users write mods, a word that needed no explanation because a generation of gamers already knew it: a mod changes how a game behaves without touching the engine underneath. Six days later, on stage with Satya Nadella, Jensen Huang said that MXC, the Windows layer that limits what an agent can touch, would do for agents what Windows and DirectX, the layer through which Windows games talk to graphics hardware, did for applications.

I suspect that is no coincidence. Games spent decades solving a problem business software has only just met: how humans and autonomous actors can share one world, and what keeps that world coherent when nobody operates every part of it. Agents now face that problem inside the company, and they keep borrowing the answers games found.

The world becomes playable

For fifty years, video games have been learning how to make virtual worlds playable. A player cannot meaningfully inhabit a world that does not respond, so games assembled a small set of primitives: GPUs, engines, simulation, autonomous actors, persistent state, multiplayer coordination, mods. Each one answers the same question: how do you make a complex world legible enough to navigate, responsive enough to act in and bounded enough to have consequences?

AI is beginning to apply that architecture to work. Gamification, the old promise, added points and badges to work that stayed the same. But what is emerging goes deeper. Organisations are building representations of their customers, factories, contracts and workflows that are complete enough for humans and agents to navigate, reason over and act upon. In parts, the company is becoming a playable world.

An ERP, a CRM or a spreadsheet stores fragments of reality. A world model connects the fragments into an environment where something can be observed, simulated and changed. Its value lies in making a complex reality operable without exposing every table, form and procedure to the person who simply needs an outcome.

The gaming stack quietly became the agentic stack

The GPU links the two at the origin. It was built to simulate graphics, the interface through which a player touched a game world. The same chips turned out to be exactly what LLMs needed, and LLMs are becoming the interface to a new abstraction of the enterprise. The sequence reads almost like a loop: render the world → model the world → act upon the world. As AI becomes agentic, it operates inside an environment it can observe and transform, beyond describing or generating.

Nvidia's own story follows that path. Founded in 1993 with video games as its way into the market, it sold GPUs to gamers for two decades; two of its GeForce gaming cards trained AlexNet in 2012, and only in mid-2020 did its data center business first overtake gaming. Gaming is now under a tenth of its sales.

Agents are the second clue. Traditional business software exposes its internals: menus, folders, forms, buttons. An agent changes the relationship. You give an entity a goal (a quest?) and let it work through the environment, coming back for judgment or authorisation when the rules require it. The nearest analogy is a player directing a group of autonomous units in a strategy game, far from a spreadsheet macro.

But the analogy runs deeper. The game engine matches the harness that runs the agent, the physics engine matches the rules and permissions that bound what it can do, the save state matches memory and provenance, and a mod matches a local change in how the harness behaves. The skills and procedures an agent can call on form a playbook: the know-how of the organisation, written down in a form that agents can run. Traditional software gives humans interfaces to databases and processes. Games give actors a world: objects they can perceive, rules they must follow, actions they can take, state they can change and consequences they can observe. Agentic software is beginning to make the same move. Games, in short, invented the right abstraction for humans and agents to inhabit the same computational world.

Research went there early. Long before companies talked about agents, labs were dropping them into game worlds to see what they would build. In 2023, researchers at Stanford and Google placed twenty-five generative agents in a Sims-like town, and a single suggestion of a Valentine's Day party turned into invitations, new acquaintances and a gathering. The same year, Voyager learned Minecraft by writing a library of reusable skills, executable code it could call on later: a playbook of its own. In 2024, Project Sid scaled the idea to as many as a thousand agents in Minecraft, who developed specialised roles, followed and changed collective rules, and passed on cultural practices. This year, Emergence World kept agents from several vendors living together for fifteen days, with tools, memory and votes on their own rules, and identical starting conditions led anywhere from stable self-government to total collapse. Games were the laboratory long before the company became the world.

Huang's comparison with Windows and DirectX points the same way. MXC, the Windows layer that limits what an agent can touch, is a stable platform that others build on, which is what graphics APIs and game engines are for.

From operator to player

For four decades, most knowledge work has made people operators of software: they open applications, move between screens and carry out procedures. In a strategy game, the player sets an intention, observes the world, allocates resources, delegates to units, accelerates time and steps in when an event calls for judgment. The knowledge worker surrounded by agents is drifting toward that role, as doing the work gives way to playing the system.

“Play” here carries its older, serious meaning: acting deliberately on a dynamic system whose other actors keep some autonomy. That is why agent interfaces increasingly borrow the visual grammar of strategy games, with status panels, task queues, maps and alerts. A good interface lets someone act within the world without operating every mechanism personally.

The idea also carries weight for organisations. A playable representation is never neutral. What it makes visible, which actions it allows, when it asks for approval and which consequences it records are all choices about authority. The physics engine of the virtual workplace is made of permissions, policies, budgets and accountability.

Why mods matter

Anthropic's Claude Code Mods, announced on October 1, give the analogy a literal form. A mod is a small TypeScript function that hooks into the agent loop. It can rewrite a prompt before it reaches the model, block or retry a tool call, approve or deny a permission request, redact a secret from an output, or replace part of the interface. Claude Code can write one on request, and the built-in /diff feature is the first to have moved out of the core and into a mod.

This follows the logic of game modding. An engine provides stable primitives, and a community changes the rules, maps and behaviours above them. The most generative ecosystems let the periphery become more inventive than the original product: Half-Life produced Counter-Strike, and Warcraft III's custom maps produced Dota.

What users ask for shifts accordingly. Once work is an agentic environment, people will want to shape how their local world behaves:

When this arrives, my agent should do that. This action needs my approval. This kind of file is handled this way.

A mod works like a local law of the virtual workplace: it changes how one particular world behaves. It also hints at a shift in software economics. Once an engine converges, innovation moves to the mods. If frontier models keep converging in raw capability, the next burst of creativity may sit in the harnesses, runtimes and communities that shape what agents do.

The real loop: virtualisation, action, consequence

The strongest version of the idea sets gamification aside. Gaming spent fifty years learning how to make virtual worlds playable, and AI is now making the real world playable through the same primitives.

The organisation first becomes a digital representation → agents and humans operate in it → their actions reach the physical and commercial world → the consequences return as new state. The digital twin of a factory grows into an operable model of the organisation itself.

The loop has a limit. A game can safely simplify its world, while an enterprise representation can hide risks, dependencies, people and power. Making the real world playable is valuable only while the model stays inspectable, the rules contestable, human judgment available and the consequences attributable. Without those conditions, the game metaphor ends up concealing control.

Work hard, play hard

There is a last irony. As per some reports, the video game industry is going through its darkest night: investors see little growth to fund, AI has pulled capital elsewhere, and AI demand has pushed up the price of the GPUs and memory that players need. The industry that spent fifty years inventing these abstractions is in crisis as the rest of software adopts them, and as its audience gains the means to build games itself. On October 7, Google released Playground, where anyone can describe a game in a prompt, play it and share it by link, with Gemini doing the building. Games are becoming user-generated, and frontier models are already learning to pick the old ones apart down to the binary.

GPUs crossed over first. Engines, persistent worlds, autonomous actors and mods followed. Work may come to look like a game for a simple reason: games discovered, decades earlier, the right abstraction for humans and agents to inhabit the same computational world.

* * *

PS: while we are asking who eats whom, hobbyists have started handing PlayStation 5 games to coding agents to turn them into native PC programs. It is early and contested, but the wink is hard to resist. Gaming is eating enterprise. And AI is (also) eating gaming.