agents-prolisted

## Overview AI agents are LLMs that act: perceiving (tools, context), reasoning (planning, reflection), and acting (tool calls) in loops until a goal is reached. The spectrum runs from fixed pipelines (predictable) through ReAct loops (flexible) to fully autonomous agents (powerful, risky). Most production value lives in the middle: structured agentic workflows with explicit control flow. This skill covers agent architecture done responsibly: the loop, tool design, planning strategies, guardrails, evaluation, and the production hardening that keeps agents useful instead of dangerous. ## When to use - Designing AI agents (support, coding, research, ops). - Choosing agent frameworks (LangGraph, CrewAI, raw loops). - Designing tools for agents. - Adding guardrails and human-in-the-loop. - Evaluating agent performance. - Debugging agent failures (loops, tool misuse). - Deciding agents vs fixed pipelines. ## Core concepts - **The agent loop.** Observe → think → act → observe... — the ReAct pattern: reasoning traces interleaved with tool calls. Everything else is elaboration on this loop. Understand it cold before adding frameworks. - **Pipelines vs agents.** Fixed sequence (deterministic, cheap, debuggable) vs dynamic tool selection (flexible, expensive, emergent failure modes). Default to pipelines; use agents where the path genuinely can't be predetermined. - **Tool design.** The agent's hands and its biggest lever: narrow scope, precise descriptions, typed inputs, informa