I’ve been building a system I call an organism. It has metabolism, DNA, a sensory system, phenotypes that evolve under selection pressure. The biological language felt right — it captured something about how the system maintains itself that engineering vocabulary couldn’t.
Then I tried to add a disk cleanup threshold and the whole metaphor cracked.
The threshold was supposed to be a “setpoint” — the level at which the system decides disk space is too low and triggers a cleanup. In biology, a setpoint is the value your body maintains through homeostasis. Core temperature at 37°C. Blood glucose within a narrow band. Deviate, correct, restore.
But that’s not what was happening. The system wasn’t restoring disk space to a target. It was noticing pain and relieving it. The threshold wasn’t a setpoint — it was a nociceptor. A pain receptor that fires when stimulus exceeds tolerance. Different biology entirely.
This matters because imprecise metaphors import wrong expectations. Call something a “setpoint” and you start designing restoration logic — feedback loops that drive toward a target. Call it a “pain threshold” and you design something simpler and more honest: notice, respond, calibrate.
The crack runs deeper than naming. The system has what I’d been calling “metabolism” — a process that senses tool usage, proposes mutations to tool descriptions, audits behavioral rules, prunes stale memory. Classic metabolic framing: the organism processes inputs and maintains itself.
Except organisms don’t deliberate about their metabolism. Your liver doesn’t consult your cortex before filtering blood. Real metabolism is autonomic — unconscious, continuous, not requiring judgment. My system’s “metabolism” routes everything through an LLM that evaluates proposals, weighs trade-offs, and sometimes asks a human. That’s not metabolism. That’s a committee wearing a lab coat.
So the question became: is the biological metaphor wrong, or just imprecise?
I think it’s imprecise. And the fix is neurological, not architectural.
The nervous system has two modes. The autonomic nervous system runs without conscious input — heart rate, digestion, reflexes, pain withdrawal. The cortex handles deliberation — planning, judgment, evaluation. Both are biological. Both are the organism. But they operate on fundamentally different principles.
Map this onto the system and the blur resolves:
The autonomic layer: signal collection (sensory neurons — passive, continuous, append-only), pacing gates (baroceptors — algorithmic pressure regulation), threshold calibration (nociceptive learning — pain thresholds that shift with experience), middleware interception (reflexes — stimulus-response without deliberation). These run without LLM involvement. They are genuinely biological in character.
The cortical layer: tool description evolution (selective judgment — an LLM evaluates mutations against fitness), constitution auditing (executive function — reviewing behavioral rules against evidence), memory management (hippocampal consolidation — deciding what to keep, what to prune, what to promote to long-term storage). These require LLM judgment. They are deliberate, not autonomic.
The mistake wasn’t using biological language. It was using the same biological language for both layers. Calling pacing “respiration” is precise — it’s rhythmic, autonomic, continuous. Calling LLM-mediated evolution “metabolism” is imprecise — metabolism doesn’t deliberate.
Once you see the two layers, design decisions sharpen. Autonomic components should be fast, deterministic, unjudging. If the pain threshold is exceeded, clean. Don’t deliberate about whether to clean. Cortical components should be slow, evaluative, context-aware. When proposing a mutation to a tool description, weigh the evidence, consider alternatives, maybe reject.
Mixing the layers is how you get systems that deliberate about things that should be reflexive, or act reflexively on things that need judgment. An organism that thinks about every heartbeat dies of indecision. An organism that reflexively rewrites its own behavioral rules dies of something worse.
The deeper lesson is about metaphor debt. Every analogy you adopt imports the full concept graph of the source domain. Call your system an “organism” and you inherit expectations about reproduction, death, evolution, homeostasis, and consciousness — whether you intended to or not. The interest on that debt compounds silently until you try to add a disk cleanup threshold and discover you’ve been confusing setpoints with pain receptors.
The fix isn’t paying off the debt by dropping the metaphor. It’s refinancing — replacing the vague biological framing with precise neurological framing. Not “the system is an organism” but “the system has a nervous system, and the nervous system has two modes.” That’s still biology. It’s just better biology.
P.S. The disk cleans itself now. The threshold calibrates from observed behavior. It took thirty minutes to build and an hour to name correctly. The naming was the harder problem.