The agent is the mind. The code is the nervous system. The robot is the body.

Start with roles, not mystique. The mind interprets a situation and proposes an action. The symbolic nervous system translates between that proposal and one particular machine. The body is the physical hardware that senses, moves, resists, and can be damaged.

This presentation uses neurosymbolic as its own architectural synthesis: neural inference in a harness-plus-model mind, joined to deterministic robot code that names perceptions, checks requests, arbitrates motion, and retains explicit artifacts. Neither source project describes itself with that term, and these roles are not a claim of sentience.

A model alone is a prediction engine. The harness gives it a working loop: collect events, construct context, ask for a response, expose a bounded tool set, validate the result, and report what happened. Together they form the agent role in this presentation.

Reachy's external `agent attach` client makes that boundary concrete. It reads the deterministic runtime feed, turns events into cues for a tool-use model, validates four intent tools, and writes atomic requests to the intent spool. It never opens the robot SDK, and its speech and pose tools are publish-only. The 50 Hz runtime keeps ticking when the client detaches.

That external client is not Reachy's folded `listen --live --cognition agent` process. Both use model-backed tool cognition, but they attach to different seams and must stay separate when we discuss extension and memory later.

Robot-specific code is the symbolic nervous system because it gives physical signals names and consequences. A direction of arrival becomes a sense event. A declarative rule can react or inhibit. An intent is checked against a catalog. Arbitration chooses one owner per motion channel. Only then does the physical hardware receive motion.

Reachy's behavior engine runs that work at 50 Hz with a passive feel-alive layer, declarative rules and modes, a durable intent spool, structured runtime events, and per-channel arbitration. The loop is deterministic and can run with zero LLM calls, so presence and reflexes do not disappear when no agent is attached.

This layer is not the neural AI and it is not merely plumbing. It is the machine-specific knowledge that makes a general reasoner grounded, reactive, inspectable, and constrained on this physical platform.

Reachy Mini's physical hardware is expressive: a movable head, antennas and body base, a microphone array, camera, and speaker. Its nervous system is organized around continuous presence—senses become events, behaviors compete for channels, and an agent can attach without taking ownership of the SDK.

The SO-101 arm's physical hardware poses a different problem. ARM101 CLI exposes agent-readable contracts around consent-gated motion, torque ownership, overload-aware movement, contact inference, append-only exploration logs, and a persisted reachability map. Its nervous system is explicit about what was commanded, what was measured, and which safety action actually succeeded.

Those mechanisms narrow risk; they do not make motion inherently safe. ARM101 infers contact from servo load and stall behavior, not from a tactile sensor, and the limitations below are part of the architecture rather than footnotes.

Use learning carefully. Here it means retained adaptation: an explicit artifact or constrained extension path that survives the event that produced it and can be inspected later. It does not mean either robot updates model weights online.

The ladder starts with persisted rules and modes, then observations and maps, then declarative behavior records, and finally generated code that must pass a validator before activation. Each rung gains flexibility by leaving a stronger review burden behind.

Nothing in this evidence demonstrates model-weight training, an ARM learned neural policy, or unconstrained self-modification. Neural inference may propose; the symbolic nervous system decides what form can be retained and what is allowed to run.

A neat diagram tempts us to draw one circle: perceive, reason, act, remember, repeat. The shipped evidence is more interesting because it contains four distinct tracks with missing connectors left visible.

Reachy's external attach track is a reactive control loop, not a learning loop. Runtime events reach a harness and model; validated intents return through a durable spool; the engine applies rules and arbitration before motion. The attached client has no forge or stash tools.

The other three tracks retain artifacts, but they are not wired together: forge auto-activation belongs to folded live cognition; stash is reachable through the Python API; and ARM exploration produces a map that current flex motion does not consume. Joining those tracks would be future integration, not a shipped autonomous loop.

Separating mind, nervous system, and physical hardware lets each layer fail differently. Reachy's deterministic runtime can stay present without neural services. An attached model proposes intents without becoming a second hardware owner. ARM101 can record measurements and report incomplete release instead of claiming safety it could not verify.

It also makes memory inspectable. Rules, spool commands, behavior records, JSONL observations, generated artifacts, and maps can be named, checked, rejected, replayed, or removed. That is a more useful claim than saying the robot simply learns.

The limits are equally important: this is an architectural comparison, not proof of sentience, a formal hybrid-reasoning algorithm, online neural training, universal contact detection, rigorous 6-DOF truth, shared code, or one autonomous closed learning loop.

The architecture rhymes, but the products do not merge. Reachy Mini CLI and ARM101 CLI target different physical hardware, expose different robot-specific nervous systems, and have no shared runtime or shipped cross-repository integration.

Follow each repository to see its current work, use the pinned evidence beside each technical beat to audit the claims made here, and visit the existing Reachy Mini agent page for the deeper single-project story. The comparison belongs to this presentation; the source projects remain independently grounded.