Agents

This Skill addresses the most challenging stability and cost issues in AI Agent development. In practice, Agents often exhibit abnormal behaviors such as infinite loops, repeatedly calling the same tool, deviating from instructions after multi-turn conversations, fabricating tool parameters, stopping tasks prematurely, or silently swallowing tool errors. Additionally, architectural choices between single-Agent and multi-Agent systems, framework comparisons, Token budget control, memory design, tool schema optimization, retry and timeout strategies, human-in-the-loop approval workflows, Eval set construction, and security hardening (e.g., prompt injection, tool abuse) are all areas where teams repeatedly encounter pitfalls when deploying Agents.

Usage is straightforward. At the start of a session, simply follow the Skill's guidance to first read ~/Clawic/data/agents/config.yaml and memory.md. It will automatically load your declared Agent configurations and historical observation records. You can then use natural language instructions to define, debug, evaluate, or harden the Agent—for example, "Add a tool to this Agent that triggers an automatic fallback after three consecutive call failures" or "Analyze the current Agent's context budget and optimize Token consumption." After each change, it automatically writes the results to the local memory file before the session ends, supporting persistent tracking.

It is ideal for teams or individuals building production-grade AI Agents, especially developers who have experienced issues like uncontrolled Agent behavior, cost overruns, or security incidents. Whether performing fine-tuning for a single Agent or orchestrating and coordinating multi-Agent systems, and regardless of whether you choose LangChain, CrewAI, or a custom framework, this Skill provides a framework-agnostic, universal methodology and practical templates. It is also suitable for QA and security engineers who need to create Agent Eval sets, regression test suites, or conduct red-team security testing.

It is recommended to use this Skill to initialize the Agent specification file before developing any new Agent and to include the memory file in version control. Note that it primarily focuses on the five core dimensions of the Agent itself: "loop-tool-memory-cost-escalation." It does not cover Retrieval-Augmented Generation (RAG), pure prompt engineering, or Agent personality settings—these scenarios require other specialized Skills. Furthermore, all read/write operations are performed locally in plain-text notes, eliminating concerns about data leakage.