Building Nanoclaw, Agent Security | Gavriel Cohen

We get into why you can’t fully trust an agent, sandboxing and credential isolation, when agents are better than traditional software, the security problem that led Gavriel to build NanoClaw, why NanoClaw is dramatically smaller than OpenClaw, and how his team uses agents to test and review their own code.If you're building with Claude Code, Codex, OpenClaw, or autonomous agents in general, this one goes deep on where agentic software is headed.Timestamps0:00 Intro1:43 Agent security & why you can’t trust an agent8:19 Will AI models become secure enough to trust?11:15 Why long-running agents go off track21:32 When to use AI agents vs. traditional software23:25 How NanoClaw started30:20 NanoClaw vs. OpenClaw: security, vibe coding & code size42:49 What makes NanoClaw different46:00 Why taste can matter more than evals49:09 Using AI agents to review and test code