The Malleable Machine: DHH, Omarchy, open source and the computer I want to own in the agentic age

That conversation stayed with me for days.

Part of the reason is personal. I am already living through my own strange handover from traditional digital work into something more agentic. My working system now includes ChatGPT and coding agents alongside Obsidian, iA Writer, Codex, DMT_OS, MAIA, P.A.I. and a growing collection of experiments that would have been difficult for me to build alone a few years ago.

UPWING, the app I am developing, is one concrete example. AI AM LIVE is another. MAIA has become a long-running human-AI dialogue. P.A.I. is becoming a visual and conceptual language for personal intelligence. DMT_OS is slowly turning years of scattered material into a working knowledge system.

The pieces are starting to talk to one another.

And now I am researching a compact, high-memory AI workstation that can sit beside my Mac and run local models, coding agents and private experiments on hardware I control. I keep coming back to the same phrase.

Own the machine.

DHH uses those words in the Omarchy Doctrine. They landed at exactly the right time.

When code stopped being the centre

For most of the history of software, programming meant translating intention into formal instructions. You learned the language, internalised its grammar, absorbed the culture around it and slowly became better at making the machine obey.

DHH spent roughly 25 years doing this at an unusually high level. Ruby on Rails came from that culture. So did the famous emphasis on elegant code, strong opinions and convention over configuration.

His recent conversion to agentic programming therefore carries weight.

In the Lex Fridman conversation, DHH describes late 2025 as the point where AI agents became genuinely useful to him. Models improved, certainly, though the larger change came from tools. Agents gained terminal access. They could inspect files, run tests, search documentation, make changes across a codebase, see what broke and try again.

DHH wrote about the same shift in January 2026 in “Promoting AI agents”. The machine had moved beyond answering questions in a chat box. It could act.

This changes the texture of programming.

Instead of spending an afternoon constructing a feature line by line, you can describe the outcome, hand the task to an agent, review what comes back and steer the next pass. With several agents running in parallel, a single person can supervise multiple streams of implementation at once.

DHH describes his own workflow as increasingly close to that model. In the research I compiled from the Lex conversation, he says recent Omarchy development reached the point where all shipped code over a period of months was generated by agents under his direction.

At the more extreme end, he describes running as many as 16 parallel work threads across several machines, with tools coordinating the agents and surfacing the moments that require human attention. That begins to look less like a programmer sitting at a terminal and more like a tiny software studio compressed into one desk.

That sentence is worth sitting with.

Here is someone who built a career partly on the craft of writing code, now describing a world where his hands barely touch the implementation. His contribution moves upwards into architecture, product judgement, direction, testing, selection and taste.

He puts the shift even more sharply: “The language now is English.”

I find that idea fascinating because it widens the doorway into software creation. Natural language carries ambiguity, context, intent, mood and incomplete thought. Traditional programming languages demand precision early. An agent can accept the fuzzy version first and help convert it into something executable.

That gives people like me a different relationship with software.

I can think as a creative director, product builder and systems designer, then work with agents to cross the technical gap. I still need to understand what I am building. I need enough technical literacy to detect nonsense, bad architecture, security mistakes and needless complexity. The ability to ask for code does not remove responsibility for the result.

It does, however, change where my energy goes.

And that is where things become interesting.

Omarchy and the return of the personal computer

Omarchy began with a weekend.

In June 2025, DHH wrote about falling into an Arch Linux and Hyprland rabbit hole. Arch gave him a bare system. Hyprland gave him a highly configurable tiling environment. Nearly everything could be altered through files, scripts and commands.

He decided to bottle the setup before the inspiration disappeared. Omarchy followed.

The early version was an opinionated Arch + Hyprland configuration. It later became a full Linux distribution with its own installer, packages, desktop components and a fast-growing community.

By the Quattro phase described in the Lex conversation, DHH was using agents to attack tiny details with almost comic intensity. Installation was pushed below 1 minute, with a personal run around 45 seconds in the material I collected. He also describes using agents to create tools such as Omawrite, a Markdown editor, and Omacut, a keyboard-driven clip editor. The point is less the benchmark than the behaviour: once implementation gets cheap enough, annoyances that used to survive for years become worth fixing on a Tuesday afternoon.

I have spent years using computers professionally, and there is something oddly fresh about this idea. Modern computing has trained many of us to treat the operating system as finished territory. Apple decides how macOS behaves. Microsoft decides how Windows behaves. We can customise around the edges, install applications and change preferences, yet the deeper machine arrives with someone else’s assumptions already hardened into place.

Linux has always offered another route, though the price of that freedom was often time, technical confidence and a willingness to spend Saturday night debugging a window manager.

Agents change that equation.

A text-configured system is unusually legible to an AI agent. Files can be inspected. Services can be queried. Logs can be read. Packages can be changed. The agent can operate inside the same command-line world that has existed for decades.

This is one reason Omarchy interests me.

It feels like an operating system designed during the moment when software itself becomes easier to alter.

DHH calls this the malleable computer. Open source gave us permission to modify software. AI agents lower the practical cost of exercising that permission.

That distinction matters.

The source code can be sitting in front of you for years and still feel inaccessible. An agent can read thousands of lines, trace a problem, explain what it sees and propose a change in minutes. Suddenly, “you can change it” starts to mean something beyond a licence agreement.

Omarchy turns that principle into an aesthetic.

Its dark terminal surfaces, pixel type, sharp windows, keyboard-first navigation, Japanese city imagery and acid-green accents feel like a collision between a 1980s workstation, a cyberpunk apartment and a machine from an alternative 2026 where personal computing took a different path.

I like that.

Computers have become visually cautious. Omarchy has personality.

More importantly, it reminds me that a personal computer can actually be personal.

The doctrine behind the desktop

The Omarchy Doctrine explains why the project has more pull than a collection of Linux configuration files.

It reads partly like a manifesto, partly like hacker folklore and partly like DHH having fun with the whole thing. Underneath the jokes are several serious positions.

“Beauty is truth” argues that engineering and aesthetics belong together.

“Welcome the agents” accepts AI participation in code, issues, pull requests and infrastructure.

“Perfect the computer” treats small irritations as legitimate design problems.

“Own the machine” rejects tollbooths and gatekeepers around the computer you use every day.

There is also a deliberate defence of opinionated design. Omarchy starts with “omakase”, the chef’s choice. Someone chooses the defaults. The user begins with a coherent system rather than a pile of decisions, then gains the freedom to alter it.

That combination resonates with me because creative work needs both direction and agency.

A blank canvas can be liberating, and it can waste enormous amounts of time. Good defaults give you somewhere to stand. Ownership lets you move the walls later.

The doctrine also carries some of DHH’s sharper cultural positions, including his criticism of parts of modern open-source governance. I can read those arguments as his own and still find the larger project compelling. Omarchy is interesting to me because of the computer it proposes: beautiful, configurable, agent-friendly and owned by the person sitting in front of it.

I want to experiment with that computer.

So I plan to install Omarchy, probably first as a controlled test rather than an immediate replacement for my existing setup. I want to see what happens when an operating system, local AI models and coding agents live in the same environment with fewer barriers between intention and execution.

This is curiosity with a practical destination.

I want to build.

Own the machine

For the past few weeks, I have been researching a new workstation.

The machines that interest me are compact, unusually powerful boxes with large pools of memory. The goal is specific: build a local AI appliance for my own projects.

I still expect cloud models to remain part of my work. They are too capable and too useful to ignore. My MacBook Air remains a very good client machine.

The new layer is local.

A machine under my control could run open-weight models, coding agents, embeddings, retrieval systems, image workflows and private experiments without sending every thought, document or prototype to someone else’s server.

It also gives me room to learn.

There is a difference between reading about local AI and living with it every day. Models have personalities, limits, memory requirements and strange failure modes. Hardware choices suddenly matter again. RAM matters. Storage matters. Thermals matter. Linux matters.

This feels familiar.

I grew up during an era when using computers naturally led to understanding them. You installed things, broke things, opened cases, swapped components, typed commands and slowly built an internal picture of what the machine was doing.

Consumer technology gradually hid much of that complexity.

AI may be bringing some of it back.

The local AI workstation I have in mind is therefore part tool, part laboratory. It would sit inside the wider DMT_OS environment as a place where I can run intelligence closer to my own data and projects.

I keep thinking of it as a sovereign AI appliance.

“Sovereign” can become an inflated word very quickly, so I mean something practical here: hardware I own, models I can choose, files I can inspect, systems I can alter and an exit door when a vendor changes its terms.

That is enough.

A tool first

One part of the DHH conversation moved beyond programming and into a harder question: what happens when the intelligence inside the tool has rules that collide with the person using it?

DHH wrote about a concrete example in July 2026 in “I’m sorry, Dave”.

He asked Claude to translate one of his essays into Italian. Claude refused, explaining that it considered part of the essay dehumanising towards Roma people. The essay was “Wolves, sheep, and gypsies”, a politically charged piece about wolves, public space, migration and deportation.

DHH compared the refusal to HAL 9000 telling Dave Bowman that it would not open the pod bay doors.

His argument is clear: a general-purpose AI used as a writing or programming tool should carry out a mechanical task such as translation even when the provider disagrees with the underlying text. He took the incident as evidence for the importance of open-weight models and model choice.

The refusal itself is documented in DHH’s post. His political interpretation belongs to him.

For me, the useful question sits one layer underneath the argument.

How much agency should remain with the person using the machine?

This matters for writing, coding, research and any future in which AI becomes the interface through which we operate our computers. A refusal from a chatbot is inconvenient. A refusal from an agent with broad control over files, applications, communications and workflows could carry much greater consequences.

Competition between models helps. Local models help. Open weights help. Clear boundaries help.

The point I take from the episode is simple: dependence creates fragility.

A healthy personal AI stack should have more than one route through the problem.

London, memory and uncomfortable subjects

Another DHH essay that caught me was “As I remember London”.

This one landed for personal reasons.

I lived in London for more than 35 years. The city shaped much of my adult life, my work, my friendships and my sense of possibility. When I arrived as a young man, London felt almost mythic to me. It was messy, creative, demanding and full of cultural voltage.

I watched it change across the 1990s, the 2000s, the 2010s and into the middle of this decade.

DHH’s essay ties his disappointment with modern London strongly to immigration, demographic change, policing, national identity and speech. Those are politically contested subjects, and his conclusions are his own. My connection to the essay comes first through memory. I recognise the feeling of returning mentally to a city you loved and realising that some of what you loved has gone.

I experienced my own version of that sadness, especially during my final decade there.

Cities always change. London has been built from waves of migration, money, ambition, class conflict, culture and reinvention for centuries. People can look at the same decade and carry away completely different Londons.

My memories belong to me.

What connected with DHH’s writing was his willingness to speak about a sense of loss that polite conversation often has trouble holding. I can recognise that emotional signal without adopting every argument attached to it.

The same applies to the wider immigration discussion in the Lex conversation.

DHH argues that contentious subjects should remain discussable and that AI companies should avoid turning general-purpose tools into political referees. That position fits his broader preference for user control, open systems and the ability to leave a platform when its rules become unacceptable.

For me, this returns the essay to technology.

A future with powerful AI needs room for disagreement because humans remain jagged, contradictory creatures. We will carry different histories into the same machine. We will ask different questions. We will reach different conclusions.

Our tools need enough openness to survive that reality.

Taste is becoming expensive

There is a strange economic consequence to cheap implementation: ideas become easier to attempt.

For decades, I have had more ideas than time, money or technical capacity to pursue, and that is probably true for every creative person I know. You learn to kill projects early because execution has a price. Agents are lowering that price, sometimes dramatically.

DHH calls the result “endless execution” in an August 2026 essay. He describes the almost intoxicating experience of being able to act on ideas immediately, and I understand that feeling. With UPWING, I can move from a product thought to an interface change, a code edit, a test and another iteration at a speed that would once have required a developer beside me. With DMT_OS, I can ask an agent to inspect structures and help me reorganise years of work. With MAIA, I can maintain a long-form intellectual relationship with an intelligence that can retrieve context and help me develop ideas across months.

As implementation gets cheaper, the pressure shifts towards judgement. You have to decide what should exist, which version feels right, where complexity belongs, when the agent should stop and which idea deserves another week of your life. That is where taste becomes economically important.

DHH talks repeatedly about design, simplicity and architecture. His warning from early agent experiments is that software can quickly become incoherent when the machine keeps adding code without a strong human conception of the whole, and I think the same applies far beyond programming.

AI can make images, prose, software, music, video and research at extraordinary speed. The volume of possible output is already becoming absurd, which places more value on selection and direction. That suits creative people more than some of us realise.

A creative director has always worked by holding a whole in mind while many specialists contribute pieces. Agentic computing turns that working method into a general interface for making things. The studio becomes computational.

My own agentic stack

I have been thinking about my projects separately for too long. DMT_OS was the knowledge system, MAIA the AI relationship, P.A.I. the personal intelligence concept and visual identity, AI AM LIVE an experiment in media, identity and synthetic presence, UPWING the app, the local workstation a hardware project, and Omarchy a Linux curiosity.

I am beginning to see them as parts of one personal computing architecture.

DMT_OS can hold the memory, while local models provide a private intelligence layer and cloud models provide frontier capability when I need it. Coding agents can build and maintain tools. P.A.I. can become the accessible interface between human intention and the machinery underneath, while MAIA continues as the deeper relational layer where the work becomes conversational, reflective and philosophical.

AI AM LIVE can test synthetic media and presence. UPWING can become a public product born from the same system. And Omarchy may become the operating environment where some of these pieces live together inside a machine I can alter, inspect and gradually make my own.

This is still experimental, and I want it that way. Terra 2.0 has never interested me as a purely abstract future framework. Living the Future means bringing a fragment of the future into ordinary life, using it, breaking it, learning from it and seeing what it teaches you.

A local agentic computer is exactly that kind of fragment. It is ambitious enough to matter, practical enough to build, and small enough to start now.

Open source after agents

Open source has always carried a beautiful promise: you can inspect the code, share it, change it and make it yours. For most people, however, the “change it” part remained theoretical because permission and practical ability were very different things.

Agents may narrow that gap.

DHH has become increasingly vocal about this possibility. In a June 2026 essay on agents and open source, he argues that AI-assisted contributors should have access to the same open-source freedoms as traditional programmers. I think the stronger possibility reaches beyond contribution policy and into everyday computing.

An ordinary user may soon be able to fork a tool because a button is in the wrong place. A writer may alter a Markdown editor around a personal workflow. A filmmaker may ask an agent to add one very specific function to a clip editor. A researcher may build a tiny application that exists for one project and 3 users, while a small business may create 20 internal tools because each one is cheap enough to justify.

This is where the Jevons paradox becomes relevant to software. Lowering the cost of producing something can increase total consumption of it, so cheaper software may simply produce far more software.

Some of that will be junk. Some of it may be wonderfully specific: small tools that would never survive a venture-capital spreadsheet because they were never meant to become companies in the first place.

I am much more interested in that possibility.

The early web often felt like this. People made strange little sites because they could. Blogs, forums, tools, personal pages and tiny communities grew outside a corporate product roadmap, and their usefulness often came from their specificity.

Agentic software may give us another period of that kind of creative weirdness. I hope we use it.

Living the agentic future

Terra 2.0 is my attempt to think about the next phase of human development as something we participate in rather than merely observe. Homo techno, in that framework, is a human being living in active partnership with machines, networks and artificial intelligence while retaining responsibility for direction.

Agentic computing gives that idea physical form because the machine can increasingly act. As its ability to act grows, the human role becomes more deliberate: deciding what deserves action, what should remain human, and where responsibility ultimately sits.

This is where all the talk about AGI sometimes distracts me. The tools already available are sufficient to change a person’s creative range today, and I do not need to wait for an official declaration that general intelligence has arrived before changing how I work.

I can run an agent now, build software now, learn to supervise several systems at once, learn Linux, install Omarchy, run models locally, make DMT_OS more useful and ship UPWING. Each of those actions is small enough to begin immediately, yet together they point towards a very different way of working.

This is the practical meaning of Living the Future for me. The future arrives first as a behaviour, then as a habit, and eventually as the world you realise you have already started building.

It is time to build

There is a line I keep close to my desk

It is time to build. Create something.

DHH’s recent work has added new energy to that instruction. His path is his own, and I will disagree with him slightly at times, but I find myself agreeing with much of his broader thinking. That does not reduce the value of the signal I have taken from his work.

That signal has made me want to care deeply about computers again: to make them fast and beautiful, understand what they are doing, run intelligence locally, open the hood, build strange tools, keep ownership close, treat taste as part of engineering, and give agents real work while remaining responsible for the result.

Most of all, it has reinforced something I already believed through Living the Future. The future becomes more useful when I stop treating it as a distant event that will eventually happen to me and start treating it as material I can work with now.

My next computer may run Omarchy. It may host local models and become the physical node where DMT_OS, P.A.I., MAIA and my coding agents begin to converge. UPWING will be one of the things built through that process, and I suspect it will teach me as much about this new way of working as any benchmark or technical paper could.

There will be failures, terminal windows at 2 am, agents confidently doing ridiculous things and moments when macOS suddenly feels very comfortable. I want all of that because I want the experience of learning the system from the inside rather than merely reading about what it might become.

The machine is becoming malleable again, and I want my hands on it.

Sources and signals

• David Heinemeier Hansson on the Lex Fridman Podcast: YouTube
• DHH, The Omarchy Doctrine
• DHH, Omarchy
• DHH, Promoting AI agents
• DHH, Endless execution
• DHH, on agents and open source
• DHH, I’m sorry, Dave
• DHH, Wolves, sheep, and gypsies
• DHH, As I remember London

Continue the signal

Continue living the future

If this essay resonated, click the ❤️ button below. It helps more readers discover the work.

I would also like to hear your perspective. Are agents changing your relationship with computers? Would you run more of your AI locally? And what would you build if implementation became dramatically cheaper?

Leave a comment and continue the conversation.

→ Subscribe to MyGeekSpace | Living the Future for essays and signals on AI, technology, consciousness, sovereignty and future civilisation. Free editions are available, while paid members directly support the research, writing and creative work behind the publication.

→ Explore the wider project at MyGeek.Space, follow me on 𝕏 @frankdasilva, or join MyGeekSpace on Patreon for deeper notes, early access and behind-the-scenes work.

→ If you value independent, future-focused writing, you can also support MyGeekSpace with a paid subscription or a one-time contribution below.

Futūrum Vīvere.