Exploring Claude Code: The Design Space of AI Agent Systems Today and Tomorrow

Claude Code is an agentic coding tool that handles shell command execution, file editing, and external service calls, and its publicly available TypeScript source is analyzed to dissect the entire architecture. At its core is a simple while-loop that calls the model, executes tools, and repeats, but the real complexity lies in the systems surrounding it.

The design unfolds around five human values and needs. human decision authority, safety and security, reliable execution, capability amplification, and contextual adaptability lead to 13 design principles, which are then connected to concrete implementations.

The main components are as follows.

• Permission system: 7 modes and an ML-based classifier
• Context management: a 5-stage compaction pipeline
• Extensibility: 4 mechanisms — MCP, plugins, skills, and hooks
• Delegation: subagent delegation using worktree isolation
• Session storage: an append-oriented storage approach

OpenClaw, presented as a point of comparison, answers the same design questions differently in a different deployment environment. While Claude Code focuses on task-level safety classification and a single CLI loop, OpenClaw takes the approach of perimeter-level access control, an embedded runtime within a gateway control plane, and gateway-wide capability registration.

Finally, based on recent empirical, structural, and policy research, 6 open design directions for future agent systems are presented.