A framework for disciplined AI coding agents
A Lotse (a maritime pilot) guides the ship through narrow waters — without taking the helm. That is exactly what this framework does with AI coding agents: it guides them from idea to merged change, without replacing the work.
The problem: AI agents drift
The pipeline
The human decides
When the shipyard builds a UI, it renders real screenshots of the result — you see what is being made, not just a description of it.
Ripe tickets arrive as readable HTML cards over the local network — pain, evidence, proposal at a glance, not a wall of text.
ready — the machine proposes, you approve.spec → ready → in-progress → merged). The status change is the membrane: the human stamps it, never the machine alone. The ledger records what was ratified — and protects it from re-litigation.The machine does the groundwork — the pen for the decision stays with the human.
The mental model — a systemic lens
»The first flies and enjoys it. The second watches the first fly. The third watches the second watching — experiencing all three levels consciously.«
inspired by »The Third Swan« (Bernd Schmid, isb Wiesloch) · our own mapping
… that is exactly what Lotse does: the same work at three altitudes at once.
Why it holds
Two-way door (small, reversible) → the bolder form is default, doing is the experiment. One-way door (data model, public interface) → full rigor before commit.
Advice becomes testable when a different model tries to refute it — not by making the advisor smarter.
Decide once, record it, don't re-litigate. A decision is a rock, not a proposal.
Machine-readable schemas + guard hooks make discipline mechanical, not just an appeal.
Every run ends with a retro about the way of working; the workshop condenses it into sharpenings. The framework is its own first use case.
What it is
commands/, agents/, contracts/, hooks/ — versioned, reviewable.Repo = truth · runtime = deploy target · the human stamps.