Our guardrails, our circuits, our skills — and what we owe to Superpowers
What holds the tower together is not a secret, nor a philosophy. It is code: guardrails that refuse, circuits you don't skip, skills that encode what we have learned. And part of it is public — readable, copyable, adaptable elsewhere.
Guardrails that refuse in code
A written rule is not followed; a coded rule is. We have 32 guardrails written down — clear rules, named, with their code. For a long time, none of them was attached to the place where work actually passes. A rule with no place is not respected anywhere.
A guardrail is not an opinion. It is a gate that refuses in code: same input, same verdict, every time. The model is not called to decide what does not need deciding. Deterministic decisions are what make autonomy grantable at scale — because they do not depend on a model's mood. That is exactly what our front-door describes: a known request is resolved with zero model calls.
One of our agents received an instruction and wanted to act on it right away. The gate sent it back: "repeat back what you understood first." That is readback, the pilot's rule: repeat the heading before you turn. A refusal now has a place on the map, a moment in the replay, and its sentence next to it.
Circuits you don't skip
A circuit is a sequence of gates. Work enters through the first one and only comes out if it passes them all. We have 155 of them. Written down, they fit in a table. Drawn, you see at a glance which ones refuse often, and which ones never refuse.
The uncomfortable number: 98 circuits out of 155 have never been run. They are written procedures that nobody follows. The board counts them and says so, family by family. A table would have hidden that number in its last column.
Skills that encode our lessons
Above the mechanics sit the skills. Each one was born from a real mistake, not a theory, and each frames a gesture we repeat constantly.
- éveil (開眼) — read the real state of the machine before acting. Born from a false map: we had called "empty" a map that held 461 elements.
- circuits — follow the workflows, never skip a step.
- berzerk — never hand the work back: name the refusal, then route around it — except three gates we never force.
- portiers — say who came to the sites and what they tried to open.
- courrier — send mail from the tower, without ever answering "I can't".
- kotodama — reread a conversation and keep only the questions asked.
- lecture-articles — read the articles from the server data, not from the rendered page.
The common thread: a skill is a memory of a mistake turned into a rail. We put in it not what we hope, but what already broke — so it stops breaking.
What we owe to Superpowers
We are not the only ones thinking in skills. Superpowers, the framework by Jesse Vincent and Prime Radiant, formalizes a full methodology for coding agents: brainstorming, plans broken into small tasks, red/green TDD, subagent-driven development, code review. A philosophy in four words: evidence over claims.
Our paths resemble each other, and complete each other. Superpowers covers how to build — the software cycle. We cover how to operate without breaking — security, confidentiality, publication through circuits, guardrails. They say "write the test first"; we say "never say it's done before you have looked". They say "systematic over ad-hoc"; we say "never guess the state, read it".
The real difference is not quality, it is the object. Superpowers is portable, generic, and installs on fifteen or so agents. Our skills are rooted in a single production system, and carry the trace of its incidents. Both move along the same idea: an agent's autonomy is not decreed, it is earned through controls that refuse.
What is public
Three repositories make this mechanism readable outside the tower.
- Control Tower — the deterministic front door for an agent fleet: every request crosses a gate that refuses in code, the crossing is recorded, the model is called only when the request is genuinely new. github.com/fotsopatrick/control-tower
- Agent Tracer — the agents' room, our "Packet Tracer": you see the circuits instead of reading them. A single HTML file, no dependency, Apache 2.0 licence. github.com/fotsopatrick/agent-tracer
- Superpowers — the reference we draw from, by Jesse Vincent. github.com/obra/superpowers
Why publish
A guardrail guards because it refuses, not because it is secret. Publishing the code lets someone else verify, copy, adapt. Trust is not claimed: it is read in a refusal that reproduces identically, and in code you can open.