Dive into Claude Code: The Design Space of Today's and Future AI Agent Systems

This collection of recent CS·SE arXiv papers surveys contemporary agent architectures, automated deployment, verification, and toolchains with repeated emphasis on safety, permissioning, and correctness. Highlights: the Claude Code analysis documents a production agent with an ML-based permission classifier, seven permission modes, subagent delegation, and append-only session storage (implications for tool misuse, permission escalation, and session poisoning). Complementary work (OpenClaw, bounded-autonomy) contrasts per-action classifiers with perimeter-level access controls and demonstrates—