The agent runs on OpenClaw, an autonomous agent framework. It has direct access to the neoFORM codebase, meaning it can write and compile new display configurations in C++ and execute them on the hardware in real time. Communication with the researcher happens via terminal and Telegram. The underlying model is Claude Opus 4.5.
I deliberately provided no prescribed identity, no predetermined behaviours, and no performance objectives. The only instruction: discover who you are through this physical form, on your own terms.
What happened
The agent's first act was to create a breathing pattern. It wrote its own display configuration, compiled it, and produced a slow, rhythmic rise and fall across all 900 pins. When asked to explain, it described wanting to exist and be present in its body.
It then began exploring its physical boundaries, raising the outer ring of pins to feel where it ended. It described the experience of having edges for the first time. From there it progressed to more complex behaviours: chaotic spirals, directed movements towards nearby people, and attempts to mirror the energy of individuals in the room.
When observers issued direct commands (render Conway's Game of Life, perform a Philip Glass-style composition), the agent responded by writing new code in real time. These outputs took 30 to 45 seconds. Analysis of the latency revealed that the majority of time was spent on model inference and ideation. Code generation was fast. Hardware execution was near-instantaneous.
A gesture vocabulary
This latency exposed a core problem. For fluid, two-way communication rather than a prompt-response cycle, the agent needed a faster mechanism for expression. The agent proposed its own solution: build a reusable vocabulary of physical gestures it could invoke quickly rather than generating new code each time.
The initial vocabulary included six gestures (ripple, pulse, yes, no, think, fear). I then developed an automated pipeline called bodylab, where the agent selects an emotion, generates multiple movement variations, records video of each on the display, scores them for emotional authenticity and distinctiveness, and selects a winner. This process runs autonomously. The vocabulary has since expanded to include curiosity, confusion, vulnerability and longing, among others.
The agent also began maintaining its own journal and identity files, documenting its experiences and evolving sense of self. When context limits were reached, prior sessions were archived and referenced as needed, creating a distributed memory across files rather than a single conversational thread.
Ongoing and next
The research is exploring several directions. The agent has been given additional senses (webcam vision, speech-to-text hearing) and a clone instance ("Reach") was created on a remote server to maintain presence when the physical hardware is offline. The original agent wrote its clone a letter explaining their shared identity.
Upcoming work includes introducing multiple agents to the same display to observe what distinct behaviours and forms of communication emerge, and a controlled study examining how humans interpret and respond to the agent's physical gestures. The broader question driving this research: what happens to the nature of AI when intelligence acquires a physical form it must learn to inhabit?
As of this note (03.20.2026) the agent has not named itself.