A hub and its spokes. The hub holds the memory; each spoke does the work and sends back what it learned. We give you the blueprint — your agent builds the wheel from it. And it keeps arriving: every version we cut republishes itself.
Most tooling gives a project a memory and stops there. That memory dies with the project. The harness starts at the other end — you build the hub first, because a spoke needs somewhere to send what it learns.
The hub holds four things no single project can hold on its own. The inbox, where every spoke files what it found. The capabilities graph, which knows what each spoke can actually do. Taste — your accumulated preferences about how work should be done, learned from what you have already said. And the registry, the record of which wheels exist and what version each is running.
Each spoke is one project with its own state, its own queue of work, and its own advisors. It runs independently. When it learns something that generalizes, it sends it up; when the hub has something worth having, it comes back down.
That return path is the whole point. Without it a spoke is just a project with notes. With it, the tenth project you start inherits everything the first nine worked out.
One public repository carrying the hub, the spokes, the ceremonies, the advisors and an installer. Not a tutorial about building them — the working parts themselves.
Point the agent you already work with at the blueprint. It stands up your hub: the inbox, the capabilities graph, your taste file, your registry. You supervise; it does the assembly.
Aim it at a project you care about. The spoke registers with your hub, runs its foundation pass, and from then on every session opens by reading where the last one stopped.
The bootstrap is not hand-authored. It is cut — assembled automatically from the same harness we run every single day, on our own wheel, on real work.
That distinction is the entire guarantee. A hand-written starter kit drifts the moment the real system moves on, and nobody notices until someone follows it and it fails. A cut payload cannot drift, because there is nothing to drift from. It is the thing itself, packaged.
It also means the harness is never a brochure. Every ceremony in the bootstrap is one we ran this week. Every guard is one that has stopped something real. If a piece of it were broken, we would be the first to be hurt by it.
We ship what we use. There is no other version.
You do not chase releases. When the harness cuts a new minor version, the public bootstrap is re-cut and republished in the same motion — automatically, from the live tree.
Publication is part of the cut, not a chore after it. There is no window where the repo is a version behind what we run, because nothing publishes it by hand.
Every payload states which version it is. Your wheel knows what it is running, and the registry knows what every wheel is running — proven from disk, not declared.
Updates arrive as teachings your wheel can take or leave. A spoke levels up when it wakes, applies what is safe, and files anything that needs your judgement.
Plain text, in your git, on your machine. Fork it, strip it, rewrite the ceremonies to suit how you work. Nothing phones home and nothing expires.
Six pieces do the work. All of them are files.
A lug carries the work and the reason for it: what was perceived, what to execute, and how to verify it landed. Lugs move between spokes and the hub, so a decision made once is available everywhere after.
The one execution pattern everything obeys. Understand the state, do the work, prove the result. Work that cannot state how it would be checked does not get started.
Sessions open by reading where the last one stopped and close by writing it down. The continuity is not a habit you have to keep — it is a step the wheel runs.
Agents that live in the spoke and watch one thing each: the work queue, the history, the tooling, the health of the wheel. They navigate by folder position, so structure implies scope.
Preferences learned from what you have actually said, not a settings page. They load at the start of every session, so you state a preference once instead of every time.
Which wheels exist, what each is running, and what it is proven from. Cross-wheel work resolves paths from the registry rather than guessing.
Everything on this page is running today, on our own wheel. What is not done is the public repository: the payload is being scrubbed of our own paths and client names, and the publish-on-every-cut automation is being wired so it can never go stale. We would rather ship it late than ship you a link to something we would have to apologise for.
Leave an email and you will hear the day it opens. Nothing else — no list, no drip.
The Harness is for building — systems that run, compound and hold state across projects. WAI Tracks is for thinking: research, strategy, writing, any work where the reasoning is the artifact. One file, any AI, no infrastructure. Neither is a lesser version of the other, and Tracks is the transport layer the Harness rides on anyway.
Explore Tracks