Paper Dives Into Claude Code: Core Is a While-Loop, Safety and Context Are the Real Work
bibryam · x · 2026-09-15
A new arXiv paper, Dive into Claude Code: The Design Space of Today's and Future AI Agent Systems, analyzes Claude Code's publicly available source code and compares it with two open-source agent systems, OpenClaw and Hermes Agent.
Key points:
- It derives 13 design principles from five human values/needs: human decision authority, safety/security/privacy, reliable execution, capability amplification, and contextual adaptability.
- The core is a simple while-loop that calls the model, runs tools, and repeats; most code lives in the surrounding systems.
- The permission system has seven modes plus an ML-based classifier; context management uses a five-layer compaction pipeline.
- Four extensibility mechanisms: MCP, plugins, skills, and hooks; plus subagent delegation/orchestration and append-oriented session storage.
- The same design questions get different answers across deployment contexts: Claude Code gates per-action safety, OpenClaw emphasizes perimeter access, Hermes renders approvals across surfaces.
Conclusion: deployment context shapes an agent system's trust model and architectural trade-offs.
More from coding & agent
- Dev builds Macross-style game iteratively with OpenAI Codex — algo_diver · 2026-09-15
- Dev builds a Macross 7 game with OpenAI's Astra inside Codex — algo_diver · 2026-09-15
- SWE: pre-nerf Opus 4.6 was the goat, and smarter models won't make SWE easier — LouMM · 2026-09-15
- A hands-on loop to learn inference engineering: vLLM in Docker, measuring TTFT and batching — malliktwts · 2026-09-15
- Cloudflare adds per-Worker access scoping and four granular roles for agents and teammates — dinasaur_404 · 2026-09-15
- Elyx debuts with a new AI-friendly design file format for designers and agents — michalmalewicz · 2026-09-15