Personal AI agents

Always-on / messaging-native agents (OpenClaw lineage, Eve, Moltworker, meta-harnesses) vs session-bound IDE agents.

Takeaways

• Persistence + tools + identity/memory beats chatbox demos.
• Security isolation is table stakes — separate machine/VM, allowlists, Docker subagents, Tailscale.
• Skills marketplaces create leverage and supply-chain risk.
• Folder-as-agent (Eve) and meta-harness (Omnigent) are competing packaging models.
• “Promoting agents” (DHH-ish): supervised terminal agents > in-editor autocomplete cosplay.

Open questions

• Where OpenClaw-style agents end and this vault’s Cursor skills begin (overlap vs complement)?
• Foundation/governance after big-lab hires.