Manifesto  ·  October 2026

A Wider Channel

When AI writes code faster than we can read it, understanding becomes the slow part. A short history of the channel between people and computers.

1. 1890
2. 1961
3. 1968
4. 1984
5. 1993
6. 2007
7. 2022
8. Now

AI can now write code faster than any person can read it. An agent can work for twenty minutes on its own and change a hundred files. The work is no longer the slow part.

So what is? Watch someone work with three agents at once. They type a request into one window and switch to another to check on a second agent. They scroll back through a long reply to find the one line that matters. When an agent finishes, they read a wall of text to learn what it did, then type a follow-up that repeats half of what they already said. The person spends most of the day moving information by hand.

This raises a question that everyone who works with AI will have to answer. When the machine writes code at the speed of light, is understanding the bottleneck? Do we need to understand what the AI executes?

Our answer is yes. You can’t steer work you don’t understand. You can’t check it, and you can’t trust it. Suppose an agent changes a hundred files and you understand three of them. You have handed over the decision along with the labor. This essay is for the people who want to shape the work: say what they want, see what the agents did, push back, and decide. For them, fast agents only help if understanding keeps up.

Bysicle exists to make understanding as fast as the work, by moving information between a person and their agents, in both directions, at the highest bandwidth we can reach.

Every era had a channel and a bottleneck

The history of computing is in large part a history of that channel. Each era gave people a new way to tell the machine what they wanted and to see what it did. Each also had a narrowest point, where a person waited, guessed, or did work by hand.

Why understanding is the bottleneck

Start with engineering. An agent can build a feature and open a pull request. Before you merge it, you need to know how it fits into the rest of the system and how you checked that it does what you asked. When something breaks a week later, you need enough understanding to find the cause. Without it, you can’t tell an agent what is wrong, and you can’t judge its guess.

The same is true outside code. In marketing, AI can make a product animation. A person still has to judge whether anyone will want to watch it. That takes a feel for pacing, music and what holds a viewer’s attention.

In sales, AI can send outreach to thousands of people. A large enterprise client needs a personal approach that shows the problem your product solves for them. To automate outreach well, you first need to know what great outreach looks like.

Three things depend on understanding. To steer, you need to know where the work is, so you can redirect it before it goes too far. To check, you need to see the assumption, so you can notice when it is wrong. To trust, you need a record of past work you understood, so you can hand over more next time.

Steering also needs a stable picture of what the agents actually did. It has to be the same picture each time you look, tied to the real files and results, so you can check it and come back to it. A picture generated afresh on every view can’t give you that, because it can change between two looks and nothing in it points back to the work.

So the job is to make understanding faster. AI can help you build that understanding too, and how you get it matters less than having it. Bysicle is one way to get it fast. Talk the work through, see it laid out, and point at the part you don’t follow. Each of these widens the channel from the machine back to the person.

Why we named it Bysicle

Steve Jobs called the computer a bicycle for the mind. That is the short form people quote, and his recorded words were close to it. In the 1990 film Memory & Imagination he described a study of how efficiently different animals move.¹⁴ A human on foot did poorly. A human on a bicycle beat every animal on the list. The computer, he said, is “the equivalent of a bicycle for our minds.” In an interview the same year for the WGBH series The Machine That Changed the World, he said, “for me, a computer has always been a bicycle of the mind.”¹⁵ The bicycle let a person’s own muscle carry them much further.

Agents add a great deal of power, and today the steering is clumsy. We named our product Bysicle, after that bicycle. It takes a rider further on their own effort, and the rider still decides where it goes. We want Bysicle to do that for a person working with agents.

The computer you already own

Bysicle is software that turns the devices you already own into an AI computer. Think of it as an operating system for language models and agents, an LLM OS. Ideally, a Mac or a DGX Spark, a pair of headphones, and connectivity for your wearables should be enough. You should not need a special device to buy and carry.

Other companies are building their own hardware and the software to run on it. We build only the software, because the devices people already have are good enough. A Mac can run speech on its own chips. A DGX Spark is built to run large models at your desk. Headphones and wearables are already on your body. What is missing is the layer that connects them to your agents and to each other, and keeps one shared context across all of them.

That layer has to be fast. We judge every design choice against one rule first: low latency, with as much of the session’s context as we can carry. A three-second pause breaks a conversation, and forgetting the last hour makes help impossible. This is why we built a native app instead of a plugin. Running speech on the Mac’s own hardware cut the time to first speech from about 700 milliseconds to about 185.

What we are building today

Bysicle today is a native Mac app. You talk to it hands-free, the way you would to a coworker at the next desk. A voice model answers in a sentence or two. When the answer is long, such as a plan or the explanation of a bug, Bysicle puts it on an infinite whiteboard as an island: a panel with tables, diagrams, code, and text. You hear the short version and see the full one.

The board works in both directions. You can point at a line in an island and ask about it, and Bysicle changes that island in place. You can draw on the board or paste an image, and it becomes part of the context. When a request needs real work, Bysicle hands it to a coding or computer agent on your machine. The agent works while you keep talking, then reports back to the same board, and you can ask what it did and why. Bysicle also keeps a knowledge base of your project and connects to your tools through MCP.

Much of this is new and some of it is rough. It runs on one platform, and it is built first for people who write software. That is where we are starting.

The plan

1. Now

   A native Mac app. Talk with your agents hands-free, see long answers as islands on a shared whiteboard, point at anything to ask, and hand off work that reports back to the same board.

2. Next

   The same boards and knowledge base on more of the devices you own. Headphones become a full way in, so you can ask your agents something and hear back from them away from the screen. A DGX Spark can run the models on your own desk. Your wearables connect as new ways to capture and ask.

3. Later

   An operating system for working with AI that runs across the devices you already have. One shared context connects what you want, what your agents do, and what gets made, wherever you are.