Summary
An agent’s coding/scripting ability is its “meta-ability” — the capacity to construct reliable tools on-the-fly rather than relying solely on neural network inference. Even non-coding agents benefit from Bash and scripting tools because they offload deterministic logic (math, data extraction, file manipulation) to reliable executors.
Key Points
- 核心洞察：编码能力 = Agent 的元能力。更强的编程能力使 Agent 能构建更可靠的”数字手脚”
- LLM 数学弱点的解法：不是让 LLM 更擅长算术，而是让它写脚本调用
awk/bc等工具做精确计算 - 数据 ETL：Bash 管道 + 正则擅长清洗非结构化文本，在 AI 处理前提取干净数据
- 可解释性优势：用户可以审计脚本代码，而非接受黑盒输出——透明决策
- 技能延迟加载（Thariq, Claude Code）：通过 Bash 工具获得”长尾用例和涌现能力”——不需要预先训练所有能力，运行时按需构建
- 不限于 Bash：同一原则适用于 Python、SQL、任何脚本语言
- 实际应用：报销计算（grep+sed+awk）、联系人去重、ffmpeg 视频处理、动态 cronjob
Relationship to Other Concepts
- Agentic Patterns: Tool use is the foundation of the “Augmented LLM” pattern — the building block for all other patterns
- Harness Engineering: Bash tools act as computational sensors (deterministic, reliable) vs inferential ones (probabilistic)
- Ashby’s Law: Tools extend the regulator’s variety without increasing neural network complexity
Open Questions
- What’s the right boundary between “let the LLM reason” vs “offload to a script”?
- How to balance tool diversity (more capabilities) with security (arbitrary code execution)?
Evidence Timeline
-
2026-04-10: “Claude Code from Source” Ch 6-7 — 14-step tool execution pipeline with self-describing tools (concurrency, permissions, rendering), speculative execution, and streaming executor.
-
2026-04-07: Created from rosa’s article “从Bash工具开始理解Agent”