Agents

This Skill addresses common loss-of-control issues in AI Agent development and operation, such as infinite loops, repeated execution of the same tool, instruction drift after multiple conversations, hallucinated tool parameters, premature task termination, and silent error swallowing. In practice, developers often spend significant time troubleshooting abnormal Agent behavior but struggle to pinpoint the root cause, especially in scenarios involving multi-turn interactions, multi-tool calls, or multi-Agent collaboration, where problems are often masked by context window inflation, runaway costs, or memory confusion.

Usage is straightforward: you simply need to ensure the local configuration file ~/Clawic/data/agents/config.yaml and the memory file memory.md exist at the start of a session. After that, you can use natural language instructions to design, debug, evaluate, and harden your Agent. It automatically reads your declared Agent specifications, observed memories, and task index, and writes any persistent outputs (such as newly defined Agents, final effective system prompts, framework selection decisions, evaluation results, cost/latency measurements, rollback packages, red team findings, etc.) to the specified locations at the end of each session. All read/write operations are confined to plain text files within the local configPaths directory—no data leaves the machine, and no credentials are ever written.

It is particularly well-suited for teams or individuals requiring fine-grained control over Agent behavior, especially those who have used frameworks like LangChain or AutoGPT but found it difficult to debug underlying loops, manage context budgets, or set cost caps and upgrade rules for Agents. For advanced scenarios like multi-Agent architectures, dynamic tool orchestration, human-in-the-loop approval, rollback strategies, and memory design, it also provides a standardized recording and evaluation framework.

It is recommended to read memory-template.md before first use to understand all write targets, formats, and thresholds. Note that it primarily focuses on debugging and hardening at the Agent behavior level; it does not cover LangChain APIs, RAG pipelines, pure prompt engineering, or Agent persona/voice customization—please use other dedicated Skills for those scenarios.