Skip to content

The Thesis

The factory that builds software — and itself.

A runtime for self-improving agentic loops where code is the first product line.

The factory is an unattended software factory. A ticket that is specified, disjoint, and agent-ready becomes a pull request without a human driving each turn. The PR merges because tests passed and a reviewer approved — never because an agent said it was done.

That is the wedge: the thing a visitor can use tomorrow, against a reward signal that already exists (CI), on a control plane that already exists (tickets), into an artifact that already exists (a pull request).

We do not lead with a general agent platform that happens to also write code. We lead with shipping software, unattended, and let generality show up as a consequence of the same primitives.

  1. The tracker holds the work: TriageTodo + ai:agent-readyIn ProgressIn ReviewDone.
  2. The worktree isolates the change: bin/worktree-up.sh provides its own branch, ports, database, and runtime home.
  3. A bounded agent implements the ticket: strictly inside its declared Owned Paths.
  4. An independent worker re-runs verification: the agent’s report is commentary; the exit code is evidence.
  5. GitHub holds the PR: CI is the reward signal. Merge and ship are later loops on the same event runtime.

The same loops that ship product ship the factory. This repository is an ordinary dispatch target. A harness defect becomes a ticket; a ticket becomes a PR; a merge changes the next run. Friction is not a retrospective slide — it is intake.

Self-improvement is not a metaphor for “the model gets smarter.” It is the factory editing its own skills, adapters, pins, and docs through the same claim → worktree → verify → PR path it uses on every other repo.

The event runtime is not a git wrapper. A research, classification, reporting, operations, or editorial agent may only need an isolated directory and a versioned input/output contract.

Those loops share the same primitives:

  • an admitted, idempotent event;
  • a content-pinned agent with a schema, timeout, and budget;
  • a deterministic planner outside the model;
  • an isolated workspace with explicit capabilities;
  • independent verification of the artifact;
  • a receipt, and optionally a chain into the next event.