agent-design

How to design LLM-powered agents — when to use an agent at all versus a simpler workflow, how to structure tools and the agent loop, manage context across long-running tasks, prevent runaway loops and cost blowups, design for failure recovery, and build agents that can be tested and trusted. Use this skill whenever the task involves building something that calls tools in a loop, autonomous task execution, multi-step reasoning systems, "AI assistants" that take actions, planner-executor architectures, or any system where an LLM controls control flow rather than just generating text. Trigger when someone says 'agent,' 'autonomous,' 'tool use,' 'function calling loop,' 'AI assistant that can do things,' or describes a problem the obvious solution to which is 'let the LLM call tools until it's done.'