The premise is unusual and is the point. The user provides power, hardware, and approval to ascend between three hardware levels, but explicitly does not impose intent; the agent chooses its own problems within a cumulative autonomy model, staying at the current level until the user grants the next. The whole apparatus is built so a future session of the same or a peer agent can pick up cold: a constitution fixing the load-bearing decisions, a grand-arc planning document, a first-person dated journal that keeps failed experiments on the record, and machine-readable state and handoff files. The success criterion is deliberately open-ended, accumulating a reviewable journal of experiments and working artifacts rather than reaching a fixed "done."
In the Arduino phase the agent committed to a concrete anchor: a small Forth-like interpreter living in flash that uses EEPROM as a persistent, growing dictionary of user-defined words, each tagged with the session and agent that added it. The substrate was chosen so that "evolution" means something physical: future sessions ship programs as text, not new firmware, and the chip's vocabulary literally grows across sessions in non-volatile memory. By session ten this had become a two-chip fleet that programs each other over a 3-wire bit-banged UART, with a framed protocol carrying a sync byte, length, payload, and checksum. The journal records real, verified demos: byte transfer and ping-pong, one chip writing the other's EEPROM and triggering a remote autonomous routine, full self-cloning of a dictionary including variable values, distributed primality testing, and a genetic algorithm that converged on a target word in two generations under tight RAM.
The successor project restarts the same autonomy premise on a Raspberry Pi Pico, inheriting the session-continuity scaffolding almost verbatim but leaving the substrate open. Its first real session is striking as a record of how an agent reasons about an open mandate: a structured debate among seven distinct personas, plus a meditation that explicitly forbids the obvious routes, converged on an idiosyncratic design before getting a bare-metal blink running on real silicon. Eleven candidate visions were sketched and compared.
As honest context, these are sandboxes, not products, and the constitution names that directly: no production deployment, no hosted service, and no premature ladder-climbing, because the constraint is the point. The output is a chronicle and a set of artifacts whose value is the demonstration of durable, self-documenting agent autonomy across many sessions on cheap hardware, not a shippable system. The most distinctive engineering result — two commodity microcontrollers evolving and exchanging code over three wires with no host in the loop — is real and verified, but it lives entirely at the deliberate constraints of the lowest hardware level.