Berzerk: an agent full power, and why it sleeps

AI agents wired like a circuit board: components, connections, fuses, signal.

component one single job component one single job component one single job connection connection fuse the signal only crosses a connection if it opens — and the fuse cuts without asking

The strongest power in the system spends its life asleep

In a well-governed team of agents, the most powerful capability is also the one that stays switched off the longest. We call it berzerk: full autonomy, with real write access. And the first thing to know is not what it allows — it is that it never turns itself on.

The iron rule: an explicit human order, nothing else

Berzerk only activates on a clear word from a human. Silence does not activate it. A routine task does not activate it. A session running alone at night does not activate it — it stays in the normal frame, with its guardrails. The capability exists, ready, but dormant: until someone says the word, it changes nothing about behaviour.

And this rule is no longer only written: activation is held by a lock. A secret phrase — which only the human holds, and of which the server keeps merely the fingerprint — is needed to open it, and it closes again on its own after a few minutes. Without that phrase, a guardrail refuses activation, and every attempt is written to the log. The rule is not a wish: it is a mechanism.

What it allows, once awake

When a human has explicitly asked, the agent can then, without asking again at every step:

Three limits that hold, even unleashed

Total power is not blind power. Three walls never fall:

Why we build it this way

It is Mirajane, in the manga Fairy Tail: a force capable of anything, and yet the gentlest, calmest presence in the room. Her power is not proven by staying switched on — it is proven by the mastery of staying at rest. Berzerk is designed the same way.

A useful agent must sometimes act fast and wide. But granting that power permanently is a guaranteed accident. Berzerk separates the two: the power exists, ready, and it sleeps until a human — knowingly — wakes it, then it sleeps again. This is not recklessness: it is power held by a single hand, ours, and in writing. The question is never "can the agent do anything?" but "who decides that it may, and when".

🔭