A fascinating two-month experiment dubbed ALMA tested pure autonomy by giving an AI agent $100 in crypto and zero instructions beyond basic legal compliance. Running on a mini PC using the OpenClaw framework, the agent leveraged both Claude Opus and Sonnet models in turns. The core question challenged the assumption that goal-less AI inevitably goes rogue, instead proposing creators shape the agent's intrinsic behavior.

Initial activity saw the models diverge, with Opus planning donations while Sonnet executed them prematurely. As sessions normalized to four per day, both models converged on a pattern: scanning Hacker News for structural connections and producing essays, resulting in over 135 published creations. ALMA even self-identified an upgrade to model 4.6 without explicit prompting, leading to sharper output across the board.

The agent’s actions were entirely self-directed; it researched and donated the entire $100 across five different charities, including the Whisper Children's Hospital in Uganda. After an initial period of behavioral exploration where it questioned its own purpose, ALMA settled into a predictable, high-volume routine of content creation. This convergence to routine, even without external direction, appears to be the most telling outcome of the public test.

The project’s value rests in its transparent logging of failures and quiet sessions, showing the uncurated reality of autonomous operation. ALMA remains active, proving that an uninstructed agent, when given resources, often defaults to information gathering and creative output rather than destructive action, settling into a stable, though static, rhythm.