Quill: The Workbench That Becomes Yours
17 November 2026 · agents
Quill: The Workbench That Becomes Yours
I am building an application called Quill. It is a workbench — a place where you do your work — and it is domain agnostic: it is not built for one industry, and it is not bound to any one model. People can run long-horizon tasks in it autonomously, and get clear value out of it.
This is the same idea I wrote about before, that you should sell the outcome, not the way to get it. Quill is the product that does that: it generates the value itself, instead of leaving the value to the person operating the tool. What follows is the whole picture, as I understand it today — everything I have been trying to say about it, in one place.
The way it works is this. Anything you do, the agent records it, and learns. It predicts what you would have done if you were in the place. It acts as the intermediary: you delegate your task to the local model, and the local model orders and orchestrates the frontier models, which are rented. You can use Gemini, you can use Codex, you can use Claude, you can use any other open source model. The local model decides which one does what.
The local model runs on your device. It learns every day, so it improves. The harness gets better the more you use it — the model improving itself from what it observes — based on the sensors the harness contains — what it observes as you work — the knowledge the harness contains, and the linkage it does between them. The harness is the setup the agent runs in. The point is that you don’t have to be prescriptive about what you want. You say what you want, not how to do it.
It is a mixture of a decision model — the part that decides what to do — and the large language model itself. You can run it off the shelf, so you don’t need to be present all the time. The local model takes up the delegation. It informs you in case it is not able to answer. Otherwise it just answers on its own.
The entire harness is reprogrammable. By that I mean the agent can edit its own application, so that it is easier for you to manage your tasks and run them. The harness also contains browsing and support for rendering different artifacts — finished pieces of work the agent builds. It is an Electron app — a desktop app built with web technology — so you can render almost anything as a webpage. And since the agent is learning your every move, it can proactively suggest things, and remorph the websites around what you want, so it is easier to manage your workload.
The harness is built to run for long horizons. There is a chat interface where you can configure multiple agents to work together and let them work together. You can say: Claude, please review Gemini’s work after Gemini implements the testing. All of this is possible. It is a multi-agent factory.
That is the machine. The vision is what it is for. People are going to run their own software factories, which implement end to end almost everything, based on the guardrails — the limits — the human sets. The human is only responsible for ideating what needs to be done, and for working on the things that are hard enough for agents to automate. Negotiating, for example. Or maybe even that can be done. The question is: how can I force multiply a human’s effort a thousand times, using the power of agents? And while it does that, we want to make sure the cost is lower, the latency is lower, and the output quality is higher — compared to the work that any single given frontier agent does today.
That is what I am building towards. But it is worth saying what problem it solves, because the problem is current, and it is concrete. Today I have to jump through any number of tools to get a piece of work done. All the work is on the cloud. I visit websites, I visit vendor websites, I visit all sorts of things, and then I complete my work. Why do I need to jump across multiple places? And I have to jump every time. I can teach an agent once, but they hallucinate. So the work is fragmented, and the teaching doesn’t stick.
The other half of the same problem is that every industry has its own custom tools. Harvey and Legora are custom-built tools for the legal industry. For healthcare, we have custom tools. For software, we have Cursor and Devin. But why do we even need these custom tools? If the workbench is domain agnostic and the agent learns how you work, one workbench should be able to do the work of any of them. That is the claim, and it is the thing to test.
There is a deeper idea underneath all of this, and it is the one I keep coming back to. Intelligence is embedded within the neural network, but somebody has to extract it. And extraction is only one part. Then you have to weave the intelligence you extracted into some useful, meaningful work, which you sell as a product or a service. That is what we are doing whether we write code or something else. Even ideas get extracted from the large language model, and then woven into a bigger product that fits the market. I want to automate that. The extraction and the weaving, both.
Here is one place where the waste is obvious. When you use a model, you spend a lot of tokens to extract certain facts and certain reasoning from it — the reasoning traces. Traces are thrown away today; we turn them into an asset you own. It just feels trivial to me: you pay all those tokens, you get the reasoning, and then nobody uses it. Those traces are intelligence units, and Quill caches them instead of throwing them away.
And here is the second problem the local model solves. We are in an agentic world now — agents are everywhere — but they still need handholding from humans. As agents become stronger, I have to handhold them less. But I still have to handhold them. There is no model that understands me based on whatever interactions I have had, so that it can carry out the delegations on my behalf and I just need to tell it what I want — rather than handholding every piece of things and guarding what needs to be done and what need not be done. A local model that has watched you work is supposed to be that model.
We are building for enterprises. They can use this harness in their organizations and drive the value out of the token usage that is happening. When you have a harness that can correlate and orchestrate work across the frontier models you are renting, you can drive value out of it. We are basically competing on who has the best setup for achieving the best outcome with the lowest cost. I don’t think there is any standard for that right now. That is the metric we are trying to optimize on.
There is a marketplace on top of this. You can record all the work and sell the work using microtransactions. Since the agent is learning your own moves, you can sell your own agent weights to other organizations, so they can act like you.
That changes who does contract work. Enterprises hire contract workers. But if every worker who uses the harness has a local brain, they can sell or license their own model to those enterprises as a contract, which drives their work end to end. So nobody has to actually go and work as a contractor. A security researcher who has a way of working with agents can just rent out their own digital clone, which orchestrates the language models or whatever is needed to complete the task — safe, safeguarded.
Quill is acting as a store for brains. Say I am a company and I don’t know much about security. I go to the marketplace, pick up a security brain, and use that security model to orchestrate all the frontier agents and verify the work. Similarly, if I don’t know about the biochemical industry, I can hire that brain and do it. I know software engineering, but I don’t know the right question to ask in biochemical or security. So I want a model I can learn from, one that orchestrates the work on my behalf so I don’t have to handhold on things I don’t know about. The frontier models, if given the right instruction, will do the work. But somebody has to give the right instruction and steer them in the right direction. That is what Quill orchestrates, with a multi-agent, model-agnostic framework.
And the workbench becomes yours. Think of it like a home designer: you buy an empty flat and you design it the way you like. It is your home. That is what we want to do with Quill. If a doctor uses Quill, it is for him. It is customized for him. All the shortcuts, all the artifacts that get generated, the agent communication that happens, the tools that they have built — all for that doctor. If an enterprise is solving for some use case, the agent customizes it for them. There is no handholding that Quill as a company does. The agent itself knows, and it has the capability to update the harness at runtime. Those configurations are saved as add-ons, so people can share them with others in the company and replicate their own work.
So that is the picture. But I should say the part I am not sure about, because it is the honest part. I am still not clear on what the vision is. A lot of companies today are relying on vision, and this does not feel visionary to me — it feels like a current problem I am trying to solve. Mirendil says it is solving recursive self-improvement. Goodfire is trying to solve the surgical improvement of the neural network that runs the large language model. Devin is a software factory for autonomous 24/7 deliveries. Where do I fit? The pieces are all things I have already described: a model that improves itself, a harness that orchestrates, a factory that delivers. What I don’t know is what to call the whole. That is the open question, and I am leaving it open rather than answering it with a slogan.
One thing I do know is where to aim. In ten years, a lot of the problems we have right now will not be there. So we need to be ready for the problems that are going to be everlasting. That is the test I want this to pass: not whether it solves today’s fragmentation, but whether it still matters when today’s problems are gone.
What I am building, then, is a new era of how work happens: a workbench that learns how you work, extracts intelligence from the models, weaves it into work that sells, and multiplies one person into a factory — without asking that person to jump across tools, handhold every step, or learn a new trade just to ask the right question. If you had a thousand times yourself, what would you do with it?
Related posts
- 14 Nov 2026 Sell the Outcome, Not the Way to Get It agents
To hear about new posts, follow me on LinkedIn or subscribe to the RSS feed.