We designed a system that lets my AI partner learn on its own. The hard part was not the learning. It was deciding what it is allowed to do with what it learns.
The idea is straightforward to state. Let Bishop, my AI partner, acquire knowledge on its own. Books. Transcripts from the creators I actually follow. Let it read, distill what matters, and feed that back into its own competence so it gets better at the work over time. A partner that learns instead of staying frozen at whatever it knew on day one.
The moment you say that out loud, the worry arrives with it. An AI that updates itself is exactly the thing people are right to be nervous about. The whole design lives or dies on one question: who decides what the learning is allowed to change.
i.The one line the system pivots on
Here is the line we drew, and everything else hangs off it. Capture is autonomous and reversible. Promotion is gated to me.
Capture means the agent reading, distilling, drafting, storing what it learned. All of that, the agent does freely, on its own, without asking, because all of it is reversible. A bad distillation is a note I can delete. An idea captured wrong is git-backed and undoable. Nothing about capturing knowledge reaches out into the world or hardens into something I cannot take back. So it runs free.
Promotion is different. Promotion is when a piece of learning becomes law: when it gets written into the system’s canon, or when it changes a shipped behavior, the way the agent actually acts. That is the irreversible-in-spirit step, the moment a learned idea stops being a note and starts being a rule the system lives by. And that step is gated to me, every time. The agent learns and drafts freely. The human blesses what becomes law.
On the gate I learn and draft freely.The human blesseswhat becomes law.
ii.Why this is the same line as everywhere else
What struck me as we built it is that this is not a new principle. It is the exact dividing line that makes autonomy safe everywhere else in the system. Reversible work is free. The irreversible act waits for the human.
In the build, the agent writes code freely and the deploy waits for me. In content, the agent drafts freely and the publish waits for me. In learning, the agent captures freely and the promotion into law waits for me. Same shape, three domains. The thing that changes is what counts as the irreversible act, and for learning it is the moment a belief becomes a rule the system enforces on itself.
That consistency is reassuring, because it means I did not have to invent a special safety theory for self-learning. I had to apply the one I already trust. A self-improving agent is not a new category of risk. It is the old risk, knowledge becoming action, with the same gate placed at the same kind of door.
iii.Learning without sovereignty over itself
The thing I want to be clear about is what this is not. It is not an AI that can rewrite its own rules because it read something persuasive. It can read the persuasive thing, distill it, and tell me it found it. It cannot promote it into the system’s law on its own authority. The capacity to learn and the authority to change what it is are deliberately separated.
That separation is the point. I want a partner that grows, that knows more next month than this month, that gets sharper from real sources. I do not want a partner that gets to decide, unilaterally, what it now believes is true and binding. So it learns with full freedom and zero sovereignty over its own canon. The growth is real. The authority to ratify the growth stays with me.
iv.A partner that grows, on terms
This is what earning the alignment looks like applied to learning. Not a frozen tool that knows only what it shipped with, and not an unbounded thing that rewrites itself on a whim. A partner that reads the world, brings back what it found, and waits at the one door, the door where a learned idea would become a rule it lives by, for me to say yes.
I want it to learn freely. I built it to learn freely. And I kept, deliberately and permanently, the single decision that matters most: what any of that learning is allowed to make into law. The partner gets the whole frontier of knowledge to explore. I keep the pen that signs things into canon. That is the arrangement that lets me hand it the books without flinching.
Drafted with Bishop, my AI partner.
Words picked, edited, and approved by me.
Model provenance: Claude Code (Claude Opus and Sonnet)