It all started, as many stories do these days, with a tweet.1This piece was written in the first week of October 2026. The field is moving fast, and some of the information may already be out of date. In the clip, a soldier from Call of Duty: Modern Warfare 2 rides a skateboard across a rooftop and lands a kickflip. The caption: "AI HACKERS HAVE PUT SKATEBOARDING IN CALL OF DUTY MODERN WARFARE 2!"
Two days later, another tweet surfaces on my timeline: in the clip, Steve, the blocky hero of Minecraft, flies in diamond armour over the royal capital of Elden Ring, the FromSoftware masterpiece.
The incursions have started.
Incursions and Mashup Gaming
My first reaction was to dismiss these clips as yet another kind of slop, aimed at teenage or nostalgic millennial gamers: franchises glued together for a few seconds of engagement. If this is slop, though, it is slop of a higher order: most of these clips show real mods, semi-working software that was vibe-coded into existence.
In some corners of TikTok, this mashup genre is named game incursions - a term coming from the Marvel multiverse, where an incursion is the moment one universe breaks into another.
The term was popularised by the modder chasm, author of the sensationalist tweet above, who also shows how a game incursion works with his SkyCraft. The mod brings the Minecraft interface and character into the world of The Elder Scrolls V: Skyrim, allowing the player to break blocks in Skyrim's world, fight its NPCs with Minecraft weapons, and invite friends into the same mirrored world.
Other examples went viral soon after: Spider-Man swinging through the Gotham of Batman: Arkham, Mario from Super Mario 64 jumping among Elden Ring's monsters, Call of Duty soldiers in Minecraft's pixelated world, and Minecraft itself dropped into Pokémon Ruby.
As with many other waves this year, it was a new LLM release by Anthropic, Claude Opus 5.5 (and its more advanced version Claude Fable 5.5), that catalyzed this wave. Expectedly, tweets about games incursions are imbued with the mellifluous tone of AI hype, and there are plenty of evangelists who are reposting these clips saying that they are realized with “one prompt”. The same rhetoric is detectable in the ambitious toolkit named Universal Modder, launched with the promise "Every game is moddable now!" and a promotional video that, as an X Community Note soon pointed out, was stitched together from other people's mods. One can't help noticing that its author works for fal.ai, a company that sells access to generative image, video, and audio models, and whose paid service the toolkit can use to generate art and sound.
But how is AI actually used in these mods? VGTimes helps us by distinguishing three methods, which differ in how deeply they enter the original game:
- The first, and most common, leaves both games intact and builds a bridge between them. Two programs run side by side on the same PC and constantly exchange data (e.g. camera position, terrain, image depth, damage), so that Steve can walk on the asphalt of Los Santos. This is how Minecraft entered GTA 5 and Skyrim, and how Spider-Man reached Gotham. A single unexpected state is enough to break the exchange, so these mods frequently crash.
- The second opens the game's code through decompilation. Game studios keep their source code proprietary, but a decompiler can reconstruct a raw approximation of it: functions without names, variables without meaning. AI Agents excel at turning it into legible code and at matching it with the original. A Nintendo 64 game estimated at 596 days of manual work has been fully decompiled this way in 84.
- The third abandons the original code altogether and rebuilds the game from scratch, with the agent doing much of the writing. The skateboarding clip comes from a complete rewrite of the 2009 Modern Warfare 2, and hobbyists rewrote the entire World of Warcraft client to add skateboarding there too.
Each of these processes used to require months of work. Former Epic Games employee Aakash Gupta noted that a studio team usually spends several quarters on a traversal system like the MW2 skate mod, while the project counted only 31 commits when he posted. The debate on the actual speed of change is interesting, but it does not alter the nature of the phenomenon. Whether or not the one-prompt propaganda holds, and allowing for a good amount of manual coding (the modder chasm says he spent weeks working by hand on the Minecraft code), agentic coding is its main technical enabler.
While I was writing this piece, the same logic left the games realm to spread into the software world at large. On GitHub, we can find an open-source toolkit ambitiously named Reverse Engineer Anything (REA) that hands decompilers to coding agents with a blunt promise: see a feature you like in an app, let the agent take it apart down to the binary, and build your own version. Another viral repository is ArtCraft, which promises to have reimplemented in the programming language Rust the entirety of the Adobe Suite (Photoshop, Lightroom, Premiere Pro, Illustrator, Acrobat).
Online commentators who tried ArtCraft were lukewarm, to say the least, about how well it actually works, and my friend and brilliant interlocutor Simone Robutti soon dropped a portmanteau for it in our group chat: slopensource.
Still, a deeper change is incubating here: what these projects share is a change in who can modify software, and under which conditions. The modders behind the incursions and the ArtCraft’s creators share a peculiar method: they work with materials already at hand, like other people's games and code, and recombine them into new objects. They point towards a figure I will call, with Claude Lévi-Strauss, the bricoleur. Its emergence reveals a contradiction: the capacity to appropriate, combine and modify software expands at the same moment as dependence on the privately owned infrastructures that make it possible.
It also raises an economic question. If a modder with an agent can rebuild a program in days, the scarcity of skilled labour that sustained the software industry's margins disappears, and the value it captured has to move elsewhere.
The Agentic Bricoleur
French anthropologist Claude Lévi-Strauss introduces his famous figure of the bricoleur in the first chapter of La pensée sauvage (1962), while discussing the technical inventiveness of non-Western peoples. From there, the anthropologist builds his famous contrast between two ways of making, engineering and bricolage. Interestingly, Lévi-Strauss traces the etymology of the latter term from the French verb bricoler, which originally applied to ball games, billiards, hunting and riding, always to describe an incidental movement, like a ball that rebounds, or the horse that swerves from its course.
If the bricoleur is, at least semantically, a figure of serendipity, the engineer starts from a project: he defines one, then procures the materials and tools it requires to execute it. On the other hand, the bricoleur works “with the means at hand” (les moyens du bord): a finite, heterogeneous stock of objects, saved because they might come in handy one day, each already shaped by earlier uses. Before starting, the bricoleur questions his stock, asking each piece what it could become. The result is therefore always a compromise between the project and the materials: the bricoleur starts from existing objects and bends them to purposes they were not designed for.
The Brazilian anthropologist and intellectual Eduardo Viveiros de Castro continues Lévi-Strauss’ work in his paper "On Models and Examples", reading this dichotomy as two poles of a continuum: every engineer becomes something of a bricoleur when his plans meet the real world, and every bricoleur is also, always, calculating and anticipating their next moves. What really separates them, he argues, is where they look for guidance. The engineer follows a model while the bricoleur is inspired by examples.
A model, in Viveiros de Castro's sense, is a general plan conceived in advance and applied to particular cases, which are expected to conform to it. It moves from the universal to the particular, in a top-down descent and it thus creates a hierarchy between those who design it (the planner) and those who must follow it (the planned).
An example is instead a particular case that inspires another particular case. It demonstrates that an operation can be performed and invites a variation on it. Examples circulate horizontally, from person to person, transforming at each passage, with no original to return to.
Models give orders and enforce order; examples give cues.
The distinction is political as much as epistemological: examples are produced by people who are forced to live inside someone else's model. The inhabitants of engineered landscapes and imposed institutions are never fully compliant: they bend and adapt elements of the model to their advantage, opening small spaces of autonomy from within. Today some of the most pervasive models we live inside are technological. As platform brutality settles in as the ambient condition, tiny examples of agency emerge from the bricoleurs online.
I first started conceptualising the bricoleur posture while preparing a lecture on cyberdecks: handmade portable computers built by non-technical people using a Raspberry Pi, the most interchangeable component in contemporary electronics, and housed in found cases, such as a vintage make-up kit or a Polly Pocket toy. When Amni Raihan asked "why are the girls building cyberdecks?", she pointed to one of the main signals of the trend's relevance: its virality, and the popularity it reached among young women and so-called "e-girls", who build and document their machines on TikTok and Instagram.
Before embracing the bricoleur as a figure of agency, we must admit that, for decades, this kind of technical subjectivity belonged to a familiar type: the white, male nerd, devoted to his tools and often resentful toward women and outsiders. Raihan points out that technical knowledge is no longer guarded by that type and it now circulates in the open, e.g. in comment sections, tutorials and playlists, where builders answer each other's questions and learn together: the circulation of examples Viveiros de Castro describes.
To this point, I would add that nowadays LLMs play a significant role in this horizontal circulation. A beginner who once had to find the right forum thread and decipher its jargon can ask a model to explain, step by step and in plain words, how to connect a screen to a board or what a terminal error means. Coding agents go further and produce examples on demand. AI does not remove the work of bricolage, which the cyberdeck scene values for its difficulty, but it lowers the threshold of entry and widens the population of potential bricoleurs.
I am suggesting that the game incursions and the emerging cases of slopensource are another instance of the same transformation, centred this time on software production. A commercial video game or proprietary software is as sealed as a smartphone: compiled, licensed and closed to every player outside the small guild of modders who spent years learning to read and modify its structure. The agent opens software once legible only to specialists, and makes it recombinable with other software worlds.
That bricolage could be a way of programming software was already argued in 1990 by Sherry Turkle and Seymour Papert in "Epistemological Pluralism". Observing children working with Logo (an educational programming language) and students in a first programming course, they found successful programmers who did not plan from the top down. These programmers negotiated with their material, tried things, adjusted them, and kept what worked. For bricoleurs, they wrote, programming is "more like a conversation than a monologue." Interestingly, LLMs have turned coding into a literal conversation: the bricoleur now programs by talking to an agent, describing a goal, testing the result, and iterating on it. From the margins of computational culture, the bricoleur style is now becoming the main entrance to software production.
Like children's play, the incursions recombine the games of our childhood into, in the words of chasm, the movement's front figure, "the games they had dreamed of as kids". One may ask whether fulfilling childhood dreams with a massive linguistic synthesiser is the way forward for our culture. It is, however, the recipe of many folk objects emerging from generative AI: the remix, an endless swap of elements made interchangeable by the tokenisation of language (the viral Hotel Lobby meme is a case in point).
This recombinant logic is precisely what destabilises the current copyright regime. The bricoleur never produces from scratch: they work by taking existing software, fragments, examples and publics, and remixing them. The biggest video game IPs have already reacted harshly towards this new wave of modding: in September Rockstar tightened its modding guidelines to forbid other companies' characters and content. Sony asked the author of a fan multiplayer mod for The Last of Us Part II to pull it before release, and Mojang had browser clones of Minecraft removed from GitHub. On 2 October, according to our modder chasm, Activision filed a complaint against his tweet showing a decompiled Modern Warfare 2.
Tolerance, on the other hand, can be part of a strategy of capture. In Brands, the sociologist Adam Arvidsson, whose work will guide us again later in this piece, argues that brands live on the productive sociality of their publics: the meanings, attachments and communities that consumers generate for free, which the brand then organises and turns into value. Fan cultures have worked this way for years, especially around pop goddesses such as Taylor Swift and One Direction, as Kaitlyn Tiffany documents in Everything I Need I Get from You. Each incursion clip, in this light, also renews the attachment to the franchises it plunders.
However, we can’t forget that the Agentic Bricoleur is playing with the master’s tools. The agent that frees them to bend other people's models runs on proprietary infrastructure, of which Claude is the clearest example, and that infrastructure can be repriced, capped or withdrawn at any moment. Here the incursions become a case study in the political economy of software and AI. And if this reminds you of feudalism, you’re not the only one.
The subscription also bundles the technical knowledge required to make the mods: the modder rents the know-how accumulated in the model. One layer above, every clip of a modded game works as an advertisement for the capacities of Opus 5.5 or a competing model. Anthropic profits directly from token usage and indirectly from every artefact its users proudly share as built with Claude. Silvio Lorusso's punchline summarises the point and could be extended: rather than vibe coding, it should be called Paid Subscription Coding.
This is the core thesis of techno-feudalism: rent becomes central to accumulation, and dependence grows on a tech elite that owns the platforms where we communicate, socialise, and now think and code. But watching the Agentic Bricoleurs scatter their vibe-coded outputs across the web, a second dynamic becomes evident alongside this centripetal, extractive one: a decentralisation of software production.
Adam Arvidsson, who already in 2019 was talking about the re-feudalization of capitalism, and Tiziano Bonini capture this ambivalence in "Factories or workhouses? Digital platforms and the re-feudalisation thesis" (2026), while correcting the diagnosis. For them, the dominance of rent does not mark a distinct return to feudalism. Phases of re-feudalisation recur throughout the history of capitalism, whenever existing social relations block the development of new productive forces. Platforms belong to such a suspended era, between the crisis of neoliberal capitalism and a new cycle of accumulation that may rest on turning social reproduction itself (communication, sociality, affects, logistics) into data and rent. Today's platforms, they argue, are the workhouses of early modern Europe. When enclosures drove masses of peasants off the land, workhouses took in the poor, housed them and put them to work, disciplining them for a mode of production that did not yet exist: the factory system.
But the workhouses did more than produce obedient workers. As economic spaces of an interregnum, digital platforms may not only incubate the most dystopian forms of life of the future, but also “radical dynamics”:
“Early modern workhouses were not simply a straightforward preparation for factory discipline in an industrial society to come. They also set in motion radical dynamics that could not be entirely controlled by the establishment that promoted them. E.P Thompson noted how workhouses – in particular the more repressive kinds that accompanied early nineteenth-century industrialisation, or ‘Bastilles’ as they came to be called – became a common enemy that contributed to unifying diverse struggles among labourers and paupers.2Thompson, 1963: 267, https://journals.sagepub.com/doi/full/10.1177/29768640261463048#bibr86-29768640261463048 At a more abstract level, the very modern individuality that workhouses contributed to fostering not only prepared for factory discipline but also paved the way for a novel politics of subjectivity that would make its mark on the twentieth century. Maybe similar radical dialectics are brewing, not just against but also within and beneath contemporary efforts and neo-feudal platformisation. Platform might not only be the testing playground for a new cycle of capital accumulation but also the (unintended) site of novel tensions, frictions and struggles between the ideological project of tech companies and the emerging ‘moral economies’ of their global users.”
The Agentic Bricoleurs may belong to these new forms of subjectivity. They already operate as nodes of a networked workshop that assembles products from shared materials, exchanging examples and savoir-faire regardless of IP law or the conventions of classic software production. If these networks organised, new decentralised economic circuits could emerge away from the major players.
Bricolage as an Economy?
The ambivalence of the current techno-feudal moment is, at bottom, a problem of value.
Tony Uphoff offers a clear account of this issue: in "Enterprise Software's Publisher Moment", he argues that software has always been, in economic terms, a publishing business. Like a book or a newspaper, a program is expensive to produce the first time and almost free to reproduce afterwards: the cost lies entirely in the first copy, while every further copy costs close to nothing. The margins of the industry therefore never depended on copying (which anyone could do) but on the difficulty of writing, on the fact that building a working program required rare skills, long hours and organised teams. Scarcity sat in the labour of writing code, and it paid twice: in the salaries of the programmers who possessed the skill, and in the margins of the software companies that organised their work. Uphoff's analogy with print is instructive in this sense: when publishing became cheap, print margins collapsed, and value moved to the aggregators that placed themselves between publishers and readers. Agentic coding, he suggests, is doing to software what the web did to newspapers. In his words: "Vibe coding is the collapse of production scarcity."
For the Agentic Bricoleur, this collapse is at once empowerment and devaluation. One person working with an agent can build what once required a studio, and for the same reason the result is almost worthless as code: any modder with the same subscription can rebuild a mod in a few days. The same logic is hitting Software-as-a-Service. The so-called SaaSpocalypse of 2026 is the crisis of companies that rent software per seat, such as Salesforce or Workday: "databases wrapped in business logic" that any agent can now easily reproduce with the right instructions. The slopensource, however sloppy, could one day threaten Adobe's dominance over professional image editing, a dominance Adobe turned long ago into a monthly subscription.
Where, then, does value move? I identify two tendencies.
The first runs upstream: to the models, to the compute that runs them, and to the materials, land and energy that keep that compute running. Training a frontier model requires hundreds of millions of dollars, a corpus on the scale of the public web, and chips that Nvidia designs, TSMC manufactures and ASML makes possible with its lithography machines. Only a handful of firms, such as Anthropic, OpenAI and Google, can finance this chain end to end, and they sell its output by subscription or by the token, at prices and under terms they set.
The second runs downstream. The authors of The Superdark Factory insist on a distinction that public debate tends to blur: a model on its own is a general capability, and it becomes useful only inside a harness, the layer of prompts, tools, context and integrations that fits it to a specific organisation, with its systems, data and habits. Building harnesses is the work of today's most hyped job title: the forward deployed engineer (FDE). The term comes from military language, where forward deployed troops are those stationed closest to the front. Palantir adopted it for the engineers it sent to work inside its first clients, defence and intelligence agencies, where they adapted its platform to each agency's data and operations on site. The work itself is familiar, since consultancies such as Accenture and Deloitte have long sold software fitted to a single client. What distinguishes the FDE is position and incentive. Rather than delivering a contracted project and leaving, the FDE is embedded in the client’s organisation and works for the joint success of the sending and the hosting company.

Value is Meaning-Making
What agents change is the cost of writing the code. I call it the last mile of software, borrowing the term from logistics, where the final stretch of a delivery, the one that reaches your door, is the most expensive and the hardest to automate. Until recently, only organisations that could afford a consultancy reached the last mile. Everyone else made do with standard software that half-fitted their needs, or with spreadsheets, paper ledgers and WhatsApp groups. Now a person with modest technical skills and an agent can build a tool around a single small business, such as the kebab shop on my street, helping to manage its suppliers and its tax deadlines. Writing that tool has become cheap, but its real cost arrives later, with maintenance: every change in regulations, suppliers or hardware requires someone who understands the code well enough to adapt it.
Here lies the value of what Clay Shirky called situated software: software built for one particular group of people and its circumstances, rather than for an anonymous mass of users, and indifferent to scale. When code costs almost nothing, what remains scarce is knowing what a program should do in a given place. As Simone Robutti pointed out to me, a program for that kebab shop is worth little as code. Its value lies in the translation it performs, from the domain of the shop to the domain of German bureaucracy. A bookkeeper who knows the Finanzamt can now turn that knowledge into a tool for an entire street, provided they know the bureaucracy before the code. Situated software fits a specific workflow, in a way that a generic, monocultural product could never. This advantage could give life to a new form of, in Yuk Hui’s terms, technodiversity: thousands of technical forms rooted in different localities and ways of life.
A glimpse of this possibility, and of its limits, comes from China. Liangzhu sits on the edge of Hangzhou, the tech capital that hosts Alibaba and DeepSeek. In a recent reportage for The New Yorker, Afra Wang describes it as the “vibe coding capital of China”. Engineers burnt out by the big metropolises have retreated there and now ship small and strange products every day, like FateTell, an app that reads your bazi, the chart of elemental forces tied to the date and hour of your birth. Behind most of these products there is no company, only a tiny team or a single person working with agents. Wang compares Liangzhu’s software market to “handmade goods from a farmers' market”. A romantic formulation, maybe, but one that captures a possible future of digital craftsmanship, in which software becomes situated by default.
In this scenario, the Agentic Bricoleur is the reversal of the forward deployment engineer: if the FDE is sent in by a vendor, the Agentic Bricoleur serves a specific locality (a neighbourhood, an industrial district, a market niche, etc.): the work Palantir sells to corporations and governments, performed for a kebab shop, a local association or a Discord server. The game incursions from which we started suggest that a similar shift could reach gaming: in a post-GTA 6 world, players might invest less in AAA blockbusters and more in experimenting with indie mashups and remixes.
Value then settles in an intermediate layer, between the labs that sell the agents and the streets that use the tools.
Adam Arvidsson, grounding his work on the historian Fernand Braudel, analysed this middle tier of the economy in Changemakers. Arvidsson argues that a middle layer of small producers, who compete within markets without dominating them, is returning today as industrious modernity: a growing number of people, many of them graduates pushed out of corporate careers, who make a living through small ventures built on shared knowledge and cheap tools. Their economy is labour intensive and capital poor, and their earnings are usually low. It is also, in Arvidsson's words, "in the midst of life". Economic calculation remains embedded in social relations, and the motivations of its actors exceed pure profit: they’re looking for creativity, dignity, and a less boring life.
The Agentic Bricoleur I’m summoning in this piece moves at this economic height: never able to set the market’s terms, but working in its interstices to capture local value.
Arvidsson, however, is the first to warn that this middle layer is never immune to capture. Capital accumulation, he argues, increasingly depends on the ability to control a multitude of industrious actors who remain formally independent of the corporations that profit from them. Digital capitalism achieves this through "a combination of mass industriousness and top–down despotic control".
His clearest example lies near Liangzhu. The "Taobao villages" are rural communities that turned into networks of small enterprises by selling through Taobao, the e-commerce platform of Alibaba, and in his account they generated 37 million jobs. Each villager owns their shop, but it is Alibaba that owns the marketplace, sets its rules and collects its data, and even regulates access to microcredit for small businesses depending on scores calculated from their online behaviour. Silvia Lindtner's Prototype Nation tells a similar story of how the Chinese state captured the maker movement born in Shenzhen in the late 2000s.
In this dialectic between capitalist centralisation and its lines of flight, what matters most for Arvidsson is the creation, or liberation, of new commons.
All the Agentic Bricoleurs we presented are slowly but certainly producing a commons of this kind: code, templates, tutorials, reverse-engineering techniques, in the form of examplesf that circulate from one bricoleur to the next. But the commons cannot stop at the software produced through these networks. The substrate on which this production runs is itself part of the political question.
The alternatives, however, exist. Agentic Bricoleurs can run open-weight models on hardware they own, groups can pool GPUs instead of renting cloud compute, and communities can maintain their own knowledge bases, model weights and templates. None of these infrastructures will necessarily match the scale or performance of the proprietary frontier. But this is not the point. Each of these choices changes where value accumulates: whether it flows upward as rent, or remains partly within the networks that produce it.
This is perhaps the most important contradiction of the Agentic Bricoleur. The same technology that makes software production more distributed can make its underlying infrastructure more concentrated. China, and Liangzhu in particular, offers a glimpse of both possibilities at once: a culture of small-scale, situated software production emerging inside an infrastructural environment that remains highly centralised.
Whether on Chasm’s Discord server or in Liangzhu, their networks are today a little of both: bazaars where people build what no publisher would have licensed, and workhouses where they rehearse their dependence on a few providers.
The Agentic Bricoleurs have learned to mod the products. The question is whether they can mod the infrastructure that produces them.
Glossary
Agentic Bricoleur. A maker who builds software by recombining materials already at hand (other people's games, code, examples) with the help of an AI agent. The figure extends Lévi-Strauss's bricoleur to agentic coding: it gains a new capacity to make, while depending on proprietary infrastructures it does not control.
Game incursion. A video game mashup in which elements of one game (characters, mechanics, interfaces) run inside another, often built with AI-assisted coding. The term circulates on TikTok and comes from the Marvel multiverse, where an incursion is the moment one universe breaks into another.
Slopensource. Open-source software produced quickly with AI agents, functional enough to circulate but too rough to rely on. Dropped by Simone Robutti in REINCANTAMENTO’s group chat.
Last mile of software. The work of fitting general technology to one specific organisation, whether by customising existing systems or by writing new code. The term comes from logistics, where the final stretch of a delivery is the most expensive and the hardest to automate.
Situated software. Software built for one particular group of people and its circumstances, rather than for an anonymous mass of users. The concept comes from Clay Shirky (2004); in this piece it names the value that remains when code becomes cheap: the knowledge of what a program should do in a given place.
Paid Subscription Coding. Silvio Lorusso's ironic name for generative AI, which I extend to vibe coding: it underlines that the capacity to code with agents is rented from the labs.
Digital workhouse. Adam Arvidsson and Tiziano Bonini's image for platforms: like early modern workhouses, they discipline people for a mode of production that does not yet exist, while setting in motion dynamics their promoters cannot control.
Upstream / downstream. The two directions in which value moves when writing code becomes cheap: upstream, towards the models, compute and energy controlled by a few labs; downstream, towards the people who adapt general technology to specific places and organizations.
--
Alessandro Y. Longo is a researcher, writer, and cultural organizer based between Berlin and Turin. He’s a PhD candidate at DREST (Italian Doctoral School of Religious Studies), where he investigates the emergence of synthetic intimacy and the problem of AI personhood. Alessandro is the initiator and maintainer of REINCANTAMENTO, an independent research and publishing group exploring technology, and radical imagination through the lens of re-enchantment. He has spent several years contributing to cooperative protocol design and blockchain governance through Curve Labs and Circles Coop.
--
I want thank my dad for sending me the first video of the game incursions and the hive mind I belong to for the valuable conversations and feedbacks, especially SR, EB, MM, MG, OSIOM, & CDO.






