Agent design lessons from Claude Code (jannesklaas.github.io)

A deep-dive into Anthropic’s Claude Code—conducted by inspecting its API with a community proxy—shows that its power comes from pragmatic simplicity rather than complex orchestration. Claude Code uses a single agent loop with a compact suite of 14 tools (bash/glob/grep/ls, file readers/writers/edits including notebook-aware parsing, web search/fetch, and TODO/task controls), plus system reminders and UX-focused features to run long, multi-step coding sessions reliably. The takeaway: you can build robust, versatile coding agents without heavyweight role-play, critics, databases, or exotic memory systems—if you design the toolset and control flow thoughtfully. Key technical lessons: Claude follows a while(tool_use) loop where model output either calls a tool (executes and feeds results back) or emits plain text (clarifying questions or final report). Work is driven by a persistent TODO tool that is created/updated as a full list (not incremental diffs) and is repeatedly injected into system reminders so the agent doesn’t drift during hundreds of steps. Sub-agents—identical Claude instances that can’t spawn further sub-agents—are used to parallelize tasks and manage context windows. For safety and speed, Anthropic routes potentially risky commands through Haiku, a smaller model that returns structured checks; humans must approve bash commands (with an “allow all similar” option). These pragmatic patterns—tool granularity, repeated in-line reminders, simple looping, and model-tiering for safety—are immediately transferable to other agentic designs.