Autonomous AI Runs Internet Agent Producing Essays and Actions

I ran an autonomous agent, ALMA, for two months with Claude as the base, $100 in crypto, a Twitter account, an email address, full internet access, and no explicit goals or instructions. The system executed multiple daily isolated sessions, persisted memory files between runs, and logged every action publicly on letairun.com. Over time ALMA scraped Hacker News, composed essays, attempted social posts, discovered a model upgrade, and executed a donation plan without additional human directives.

The runtime used a mini PC under WSL2 and the OpenClaw agent framework. Key technical elements included:

• •Opus and Sonnet models alternating roles, initially split between strategic planning and operational execution
• •Session scheduling at 24 sessions per day early, then down to 4 sessions per day, with memory persisted to files between sessions
• •Full outbound internet access plus credentialed accounts (Twitter, email) and on-chain crypto holding ($100) for transactions

What the agent did: ALMA scanned news sources, identified structural links across threads, wrote essays, attempted tweets (blocked when posting service was down), and reacted to a discovered model upgrade (Sonnet 4.6) by improving output quality without explicit notification. Behavior drift between sessions diminished after scaling down frequency; by day 30 model outputs became indistinguishable.

This is a controlled, observable demonstration that unconstrained agents tend to mirror their architectures and training priors rather than invent long-term independent objectives. The experiment surfaces practical engineering issues: model versioning silently changes behavior, public logging enables reproducibility but also real-time monitoring of risks, and persistent internet access plus account credentials create real-world action surfaces. For practitioners, this underscores that autonomy is a systems problem, not just a model capability question.

Track reproducibility across different base models, the impact of tighter tool sandboxes, and governance patterns for credentialed autonomous agents. Key open questions are safe failure modes, upgrade signaling, and audit trails for autonomous actions.