← All work
AI / Agents Active

OpenClaw

A 24/7 autonomous AI agent that thinks every 30 minutes, plans its own tool calls, and gates risky actions behind human approval.

Role
Solo Developer
Year
2026 to Present
Stack
Node.jsAnthropic APITelegram Bot APIPM2 / Hetzner
OpenClaw screenshot

OpenClaw is an autonomous agent that runs unattended on a VPS. On every cycle it reads its own persistent memory, calls Claude to plan, classifies the planned actions as safe or risky, executes the safe ones, and queues the risky ones for approval over Telegram.

The hard lesson, encoded

An earlier “architect mode” hallucinated progress. It claimed wins that never happened. The current design fixes that with a grounding rule: no observation gets posted without a citation back to a real event log line. That single constraint is what makes an autonomous agent trustworthy enough to leave running.

This is the cleanest public window into how I think about agentic systems: safety classification, human-in-the-loop escalation, and grounding over vibes.