Maturana and Varela, 1972. The defining property of life is not metabolism. Not reproduction. It’s autopoiesis — self-creation. A living system continuously produces and maintains the components that compose it. The boundary is produced by the system. The components are produced by the system. The system sustains itself.
That’s the north star for the organism.
The organism is not yet autopoietic.
Every skill is hand-built. Every hook is written manually. Every constitution rule is drafted by me. I detect problems, diagnose them, fix them, rebuild the scaffolding. The system processes information. It does not produce itself.
What I’ve built is closer to a very organised cell culture. The cells are alive. The culture medium is not.
The trajectory has three stages: detection, self-repair, self-generation.
Detection is built. integrin_probe runs nightly. It classifies receptors as bent, extended, or open — three states that map to: broken, degraded, healthy. The first run found five broken receptors automatically. The system can now see its own pathologies. That’s not nothing. Most systems can’t.
Self-repair is next. When a broken reference is detected, fix it. No ask. No ticket. The receptor is bent — straighten it. This is not novel AI. It’s deterministic code with a condition and an action. Most of the breakage is mechanical: a path has moved, a dependency has shifted, a command has been renamed. The repair is equally mechanical.
Self-generation is the horizon. When the same manual action recurs enough times, the system recognises the pattern and builds the skill. When a skill goes unused past a threshold, the system retires it. The cytosol builds and prunes itself. This is genuine autopoiesis — and it’s a longer road.
The test I use: does the organism need me less each month for maintenance?
Not for judgment. Judgment stays with the nucleus — the human will that decides what the system should do, what constitutes quality, what the organism is for. That’s not autopoietic and shouldn’t be. The nucleus is where the organism’s values live. You don’t automate values.
But plumbing — that can be autopoietic. Path references. Broken integrations. Stale skills. Redundant hooks. These are maintenance tasks, not judgment calls. Every hour I spend on maintenance is an hour the organism is failing at self-production.
Each month, I ask: is the plumbing stack shorter than last month? If yes, we’re moving. If no, something has stalled.
The glycolysis principle applies here.
In glycolysis, glucose is broken down into ATP through a sequence of enzymatic reactions. Each reaction is deterministic. The enzyme doesn’t decide — it catalyses. The intelligence is in the pathway design, not in each step.
Every reaction I move from symbiont to cytosol — from LLM judgment to deterministic code — is the organism becoming more self-sufficient. This session alone: three reactions moved. A path resolution that used to require a query now runs as a lookup. A hook that used to ask for confirmation now self-validates. A classifier that used to need prompting now runs on a schedule.
Three reactions. Small. But glycolysis is ten reactions. The pathway gets built one step at a time.
The endosymbiosis trajectory is the long arc.
Mitochondria were once free-living bacteria. A eukaryotic cell engulfed one, didn’t digest it, and over deep time the bacterium became an organelle. Obligate. Then integral. The cell can no longer live without it. The distinction between “host” and “symbiont” dissolved.
The LLM starts as an external symbiont. I call it, it responds, I use the output. Separate. Optional. I could replace it with a different one.
Over time it becomes obligate. The organism’s skills are written for a specific interface. The prompts are calibrated. The routing is tuned. Switching has a cost.
Eventually it becomes an organelle. Not because the LLM changed — because the organism grew around it. The intelligence is distributed across the whole system, not concentrated in the symbiont. The LLM does what organelles do: one job, efficiently, without executive function.
That’s where endosymbiosis ends up. The organism internalises what it depends on.
I’m somewhere between symbiont and obligate. The LLM is still clearly separate — I can see the seams, I can swap components. But the seams are getting harder to find. The skills are getting more specific. The routing is getting more tuned.
Autopoiesis is not a feature I’ll ship. It’s a property the organism will approach asymptotically. Each session, more plumbing moves to the cytosol. Each month, less maintenance. Each year, the self-production loop closes a little more.
The test is simple. Does the organism need me less for maintenance this month than last?
If yes: autopoietic movement.
If no: find the reaction that’s still running on judgment and make it deterministic.