6

Source document

[2026-03-15 03:03] traves_theberge: The key insight: LLM + loop + tools = agent The Agent Loop The core while-loop Code: basic loop skeleton Stop conditions: end_turn, max_iterations, human approval Sampling (The Model Layer) Making API calls Parameters: model, max_tokens, temperature, system Stop reasons: end_turn, tool_use, max_tokens, stop_sequence The Tool System Tool definition (name, description, input_schema) Tool registry Tool execution Returning tool results Context Management The messages array System prompts Context window limits Compaction strategies State & Persistence What state to track across turns and checkpointing strategies Orchestration Patterns Single agent, orchestrator with subagents, and parallel execution approaches Human-in-the-Loop When to request approval, handling interrupts, and streaming responses Error Handling Retry logic, tool failures, and managing context overflow Building Claude Code Integrating everything with file system tools, bash execution, and code understanding [2026-03-22 05:51] traves_theberge: (files: image0.jpg) [2026-03-23 02:21] traves_theberge: https://github.com/shanraisshan/claude-code-best-practice [2026-03-31 20:54] lisamegawatts: https://x.com/altryne/status/2039005970865516955?s=20