The body blueprint in AI
When people call LLMs "brains," they compress too much. The brain has the hippocampus, prefrontal cortex, amygdala, hypothalamus, and cerebellum. An LLM maps to one subsystem: the association cortex.
It does not store episodic memories, weigh options against goals, or respond to internal drives. Those belong to the hippocampus, prefrontal cortex, and limbic structures - subsystems absent from a transformer trained on text.
The surprise of the last five years is how far the language subsystem alone gets you.
Text is a lossy compression of human experience. At internet scale it carries enough structure that a next-token predictor recovers a shallow world model. The model knows nothing of the body, but the text corpus was written by bodies.
The harness is the rest of the brain
A harness is the control loop around an LLM, managing planning, tools, and memory. The ReAct loop is the prefrontal cortex: it maintains a goal across steps and inhibits the model's next-token drift. It decides when to stop planning and act.
Tool selection is the basal ganglia. When the harness routes among search, code execution, or file I/O, it disinhibits the tool with the highest expected reward given the context.
Memory retrieval maps to the hippocampus. Vector databases handle semantic association; event logs with timestamps handle episodic binding.
The hypothalamus analog is the system prompt. It is not reasoning; it is a persistent set-point that constrains all downstream behavior. Safety constraints limit what the model can output the way body temperature limits what the cortex can command.
The cerebellum analog is the structured output parser. The cerebellum does not generate movement; it fine-tunes it. The parser does not generate tool calls; it validates and retries them.
The nervous system is the communication fabric
The nervous system is the signaling infrastructure that lets every component communicate. The spinal cord's reflex arcs are input validation and circuit breakers that handle routine deviations without consulting the model.
The peripheral nervous system splits into afferent and efferent divisions. Afferent pathways are web scrapers, file readers, API responses, user keystrokes. Efferent pathways are tool calls, terminal commands, file writes, message outputs.
The autonomic nervous system runs infrastructure the system needs kept stable. Rate limiting and failover to backup models are sympathetic responses to load. Log compaction between turns and connection pool maintenance are parasympathetic functions.
The enteric nervous system maps to the database layer. Its 500 million neurons process, index, and retrieve without the model knowing how. It sends signals upstream that constrain behavior, the way hunger and satiety shape decisions.
Biology evolved one protocol stack and conserved it across 500 million years. Engineers rebuild it from scratch for every system: HTTP, gRPC, MCP. The design works, but the integration is rebuilt each time.
Body, hands, and feet are the execution surface
The body is the persistent container that gives the system ongoing existence. The compute instance, filesystem, and process table maintain state across invocations. They impose resource ceilings the way metabolic limits constrain the brain.
The body absorbs failure. A process crash is a broken bone; a kernel panic is cardiac arrest. Both trigger recovery procedures that assume the substrate can be restarted.
Hands are creation effectors: file writers, database mutators, PR creators, and deployment triggers. A raw POST request is a muscle twitch, but a tool with response validation and structured retry is a hand.
Feet are navigation effectors: coarse, whole-system movements like directory changes, repository cloning, SSH into another host, container orchestration, browser navigation. The model decides where to go; the feet handle the mechanics of getting there.
The Pattern
The same functional divisions emerge because they solve the same architectural problems. A fast associative core needs deliberate control infrastructure. That infrastructure needs a communication fabric and effectors at different scales.
The gap is unification. We assemble AI systems from parts never designed to fit together. The body blueprint is a functional checklist.