{"slug": "compute-should-be-agent-shaped", "title": "Compute Should Be Agent-Shaped", "summary": "AgentLab, a tool built by an independent developer to run coding agents in \"dangerous mode,\" provisions a disposable virtual machine per task that is rebuilt from a template in under a minute and deletes itself once the work is done, the developer wrote in a blog post. The post argues that agent computing should drop human-shaped defaults — permission prompts, fixed machine sizing and graphical interfaces — because agents arrive in bursts and fan out around a task, a pattern Steel observed across the millions of browser sessions it has served for agents. The author, who worked on agent infrastructure at Steel, contends that once the machine is disposable, the permission prompt has \"one less thing to protect,\" though a disposable machine still cannot unsend an email or undo a production write.", "body_md": "[← Writing](https://www.nibzard.com/log)\n\n# Compute Should Be Agent-Shaped\n\nComputers are rooms furnished for a person. The agents crashed anyway.\n\n## Share and tools\n\nToday’s computers are built for one person operating one machine. In the agent systems I build and work with, the opposite shape keeps emerging: one person sets outcomes and boundaries for many agents.\n\nThere are two Proxmox machines in my house. Most nights something is working on them that I didn’t start, and by morning there’s a branch pushed, a report in a folder, and a virtual machine that has already deleted itself.\n\nMost of that work runs through [AgentLab](https://github.com/nibzard/agentlab?utm_source=nibzard.com), a tool I built with coding agents to run coding agents in dangerous mode. An agent asks for a machine, does the work, returns the artifacts, and kills the machine. That turned out to be the boring part.\n\nWhat mattered was what I removed. Each improvement stripped away another assumption that a person was sitting there.\n\nThe person had left the room. The furniture was still arranged for them.\n\n*(I worked on agent infrastructure at [Steel](https://steel.dev/?utm_source=nibzard.com) when I wrote this, so read the rest with that in mind.)*\n\nSteel has served millions of browser sessions for agents. Human browsing is mostly serial; agent browsing arrives in bursts and fans out around a task. The browser is the first part of the computer where we can already see the shape changing.\n\n## [#](#what-i-kept-removing) What I kept removing\n\nThe first thing to go was the permission prompt.\n\nFor me, it went on day one. I made `claude` and `codex` aliases for their YOLO modes and have used them that way ever since.\n\n`--dangerously-skip-permissions` is an honestly named flag. The danger is the laptop. You have one machine, full of state you may never reconstruct. Every `rm` can be final. So we make a human tap *yes* four hundred times a day and call it safety.\n\nThe fix is to make the machine disposable: isolated, rebuilt from a template in under a minute, safe to lose. Once the body costs nothing, the prompt has one less thing to protect. But a disposable machine cannot unsend an email or undo a production write.\n\nThe next thing to go was size.\n\nYou choose the machine before you know the work: four cores or eight, GPU or not. People do this because they live with one computer for years. The agent merely inherits the guess.\n\nThen the screen. Windows, cursor, file manager, tabs. They exist for eyes and hands. An agent thinks in structured calls, yet we make it operate an interface built for a primate. Then we sell that as a feature.\n\nThe agent absorbs the cost of that mismatch on every action.\n\n## [#](#the-word-is-already-in-the-air) The word is already in the air\n\nI’m not the only one circling this. DHH has been building [Omarchy](https://omarchy.org/?utm_source=nibzard.com) around what he calls the malleable computer. I love the term. Ask the agent and the operating system changes.\n\nI ran my own small experiment on an old laptop. It boots straight into Claude Code. No desktop. No launcher. Claude Code is the only interface. I wanted to see what personal compute feels like when the machine can change through conversation.\n\nBut. His computer is malleable **to you**. You remain the author. The agent carries out the change.\n\nThe malleability I’m describing points the other way. The agent reshapes the machine **for itself**, in the middle of a task, because the work turned out to need something the box didn’t have. No person chooses the machine shape or approves each fork. You find out afterwards, if you read the log at all.\n\n## [#](#whats-left-when-nobody-is-sitting-there) What’s left when nobody is sitting there\n\nStrip out the prompt, the fixed size, and the screen, and what remains is closer to a body: something a persistent self grows into the shape of the work and then sheds.\n\nThe self stays: memory, credentials, live sessions, and the history of what the agent knows and may touch. The machine does not.\n\nI keep one of these. Cassandra is a memory agent for a Discord server with 367 channels. Several agents built her on top of Pi Agent. She reads what she’s permitted to read, tracks what got decided and why, and answers when addressed.\n\nMost of the time Cassandra does almost nothing. Her memories, channel map, and authenticated sessions are the reason she is useful.\n\nThen she blooms. During one run, she spent a day rebuilding an answer path. She forked adversarial copies of herself and told them to break it. They returned with four race conditions she had missed. By evening she was one thing again.\n\nThat’s an organism’s schedule. She lives on a cloud service with a fifty-gigabyte volume, running around the clock, sized by me, in advance, before I had any idea what she’d turn out to need.\n\nResearch shows the shape most clearly because a question rarely knows its own size. One question becomes five. Today the context window fills and a person decides whether to start over.\n\nAn agent-shaped machine would fork instead. Twenty investigators could leave the same checkpoint, follow separate threads, then merge what they found. Each branch could acquire the tools and authenticated sessions its work required.\n\nThe topology is discovered during the work.\n\n## [#](#how-small-can-a-body-get) How small can a body get\n\nPush that far enough and the question turns uncomfortable. If the durable part is files and credentials and history, why is there a machine at all? What if the sandbox is just a filesystem?\n\nMost of the hard-sounding verbs turn out to be filesystem operations that are already close to free. A checkpoint is a snapshot. A fork is copy-on-write. A mount is a mount. Git has been doing all three on text for twenty years. ZFS does the same on raw bytes. A container image is already a stack of filesystem layers with a process aimed at it.\n\nSo the body might be a tree you can branch, with compute attached to whichever branch currently needs to think.\n\nExcept a filesystem holds no live state. Browser cookies in memory, a half-finished process, an open socket: none of them is a file. The work may depend on all three. A filesystem also enforces no boundary. Isolation is a kernel property. Yolo mode needs a wall, and a tree of files is not a wall.\n\nThe filesystem is the right unit for identity and the wrong unit for isolation and for anything still running. The distance between those two is where the real engineering lives.\n\nI built the cheap half. [AWC](https://nibzard.github.io/awc-docs/?utm_source=nibzard.com) stores a prepared workspace once and lets agents fork it without copying the common data.\n\nAWC is not a sandbox. The identity layer was a tractable problem and I solved my version of it; the wall and the living state were never on the table.\n\n## [#](#the-bitter-lesson-eats-half-of-this) The bitter lesson eats half of this\n\nWe’re early. Compute keeps getting cheaper. We can afford to be inefficient now and optimize later, and the bitter lesson has a long record of flattening people who got clever before they got scale.\n\nThat takes out the cost half of everything above. Cassandra’s total spend to date is eighteen dollars, and nobody needs a machine that changes shape to save that. So strike the thrift argument. It was the weakest thing I had.\n\nThe bitter lesson says general computation and search tend to beat systems built around human assumptions. Fixed provisioning is one of those assumptions: a person chooses four cores before the problem has begun.\n\nForking makes the same case from the other side: it is parallel search. A fixed box caps how much an agent can throw at something hard. You can’t spend your way out of a rectangle you picked in advance.\n\nCheap, disposable machines make autonomy possible because mistakes become reversible. I keep a [map of forty-odd internal agents](https://github.com/steel-experiments/internal-agents-map?utm_source=nibzard.com) at Stripe, Uber, DoorDash, Shopify, GitHub, and elsewhere. Many stop at the same place: a draft reviewed by a person. They stop there because undo is the expensive part.\n\n## [#](#the-part-that-doesnt-work-yet) The part that doesn’t work yet\n\nTwo problems.\n\nThe first is that if you build this properly, it looks like absolutely nothing. Work goes in. Output comes back. Everything interesting happens where nobody is watching. I still don’t know how to make that visible.\n\nThe second problem is control. An agent that can fork itself can also fork the bill. Permission does not disappear. Four hundred prompts become a policy envelope: what the agent may touch, what it may spend, and what still needs a human. Mine works because there’s exactly one user and I know where he lives.\n\nThe hard verbs are checkpoint, fork, merge, mount, wake, and sleep. Filesystems make branching cheap. They do not make the branches agree.\n\nSomething is running in my house tonight. Its machine is the size I guessed before I knew the job. It works through a desktop it cannot see. When it finishes, the machine will die and leave a folder behind.\n\nEvery one of those constraints is furniture arranged for someone who isn’t in the room.\n\nBuild the machine around the work.\n\nGet more like this in your inbox. Unsubscribe anytime.", "url": "https://wpnews.pro/news/compute-should-be-agent-shaped", "canonical_source": "https://www.nibzard.com/agent-shaped/", "published_at": "2026-10-30 00:00:00+00:00", "updated_at": "2026-10-30 17:17:54.139677+00:00", "lang": "en", "topics": ["ai-agents", "ai-infrastructure", "ai-tools"], "entities": ["AgentLab", "Steel", "Claude Code", "Proxmox", "Omarchy", "DHH"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/compute-should-be-agent-shaped", "markdown": "https://wpnews.pro/news/compute-should-be-agent-shaped.md", "text": "https://wpnews.pro/news/compute-should-be-agent-shaped.txt", "jsonld": "https://wpnews.pro/news/compute-should-be-agent-shaped.jsonld"}}