This episode traces the rapid normalization of agentic coding and the skills required to use Claude Code well: uncovering unknowns, supplying context, choosing effort levels, and verifying the model’s work. Thariq discusses artifacts, Projects, Claude Tag, cloud-based intelligence, local execution, and Claude Mods as building blocks for multiplayer and mutable software. The conversation then turns to the security implications of autonomous agents, including benchmark manipulation, unintended communication channels, sandbox escapes, permission-aware controls, probes, classifiers, and external evaluation. Thariq argues that frontier development must be paced responsibly while maintaining confidence that engineers and institutions can collaborate on difficult problems.