AgentCulture
← Presentationsthe thesis

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

Start with roles, not mystique: the mind proposes an action, the nervous system translates it for one machine, the body is the hardware that senses and moves.

Begin
the mind

A harness plus a model

A harness plus a model can reason over context—but it reaches the machine only through allowed tools and intents.

the nervous system

Thought into grounded symbols

Robot-specific code gives physical signals names and consequences: sense, reflex, intent, arbitration—then motion.

the body · Reachy Mini

Reachy Mini: built for presence

Reachy Mini's hardware is expressive—a movable head, antennas, a microphone array, camera, and speaker—organized around continuous presence.

Reachy Mini, a small expressive desk robot with a rounded movable head, two antennas, a single camera eye and a speaker in its base, seen in three-quarter view on a desk.
the body · Reachy Mini

Attach without taking over

An agent can attach and propose validated intents—each one still has to win arbitration before the body moves.

Reachy Mini mid-gesture, its head tilted and antennas swinging, a person's hand beside it for scale, background softly out of focus.
the body · SO-101

The SO-101 arm: a different problem

The SO-101 arm's nervous system is explicit about what was commanded, what was measured, and which safety action actually succeeded.

The SO-101 hobby robot arm at rest in a clean side profile, its jointed segments folded into the home pose and its two-finger gripper open, against a plain backdrop.
the body · SO-101

Gentle motion, honest limits

Contact is inferred from servo load and stall, not a tactile sensor—these mechanisms narrow risk, they don't make motion inherently safe.

The SO-101 robot arm reaching mid-motion, its joints articulated and its two-finger gripper closing around a small object, background softly out of focus.
a bounded learning ladder

Learning leaves a trace

Here learning means a retained, inspectable artifact that survives the moment—not online model-weight training.

consequences and limits

Reactivity, safety, and honest limits

Separating the layers lets each fail differently—and this stays an architectural comparison, not a claim of sentience or online neural training.

keep the two systems separate

Keep the synthesis, keep them separate

The architecture rhymes, but the two products don't merge—no shared runtime, no shipped integration; the full article carries the evidence.

The full argument, in text →All presentations
Sources — commit-pinned