I design how people live with long-running AI agents: when the agent should speak up, how it reports, how its owner stays in control without babysitting, and how it changes without drifting. My proof is a persistent multi-agent system on Claude, Codex, Grok and Gemini models that has run research and publishing for penchan.co since January 2026. I design and run it; the models write most of the code.
Newest: OldFriend, an open-source plugin that gives the assistant you already use memory and corrections, so it grows with you.
Start here: OldFriend · auto-optimization · agent-self-healing · AI agent memory guide
- Speaking up. Health checks notify only when they find a problem; scheduled reports and a monthly letter summarize the rest. Watchdogs check whether the agent kept its promise, not only that it runs.
- Control without babysitting. Scripts do routine upkeep, hourly to monthly; a model only makes judgment calls. A daily audit flags orphaned files and broken @imports. One model family reviews another's code.
- Change without drift. A nightly dreaming loop logs events that contradict the agent's self-picture; weekly and monthly passes propose revisions. Core rules change only with my approval; self-model lines need a 30-day counter-evidence window and a check by another model family, and I hold a veto.
- penchan.co: AI and digital-finance education in four languages; 14K+ followers across platforms as @p3nchan.
- Design: former Chief Marketing & Design Officer at DA Capital, where I authored its brand and design guidelines.
- editpen: a linter that flags translated- and AI-sounding Chinese for writers. MIT code, on npm.
- CS master's student at NCCU; thesis on LLM citation verification.
- OpenClaw contributor (PR #50500).
- cc-orchestrator: Claude Code orchestrating Codex workers.
- prompt-shielder: catches tampering with agent config files.
- multi-agent-patterns: which patterns to build, defer or kill.
- More writing: agent memory explained
X @p3nchan · penchan.co · About me