ai-agents

name	ai-agents
description	AI agent architecture, tool use, memory systems, multi-agent orchestration, and safety patterns
layer	domain
category	ai-ml
triggers	["ai agent","agent architecture","tool use","agent memory","multi-agent","agent orchestration","agentic workflow","ReAct pattern"]
inputs	[{"objective":"What the agent should accomplish"},{"tools":"Available tools and APIs the agent can use"},{"constraints":"Safety boundaries, cost limits, human-in-the-loop requirements"},{"architecture":"Single agent | multi-agent | hierarchical"}]
outputs	[{"agent_design":"Agent architecture with tool definitions and memory strategy"},{"tool_definitions":"Tool schemas with input/output specifications"},{"orchestration_logic":"How multiple agents coordinate"},{"safety_guardrails":"Input/output validation and boundary enforcement"}]
linksTo	["prompt-engineering","rag","error-handling","logging"]
linkedFrom	["cook","plan","research"]
preferredNextSkills	["prompt-engineering","rag"]
fallbackSkills	["sequential-thinking"]
riskLevel	medium
memoryReadPolicy	selective
memoryWritePolicy	selective
sideEffects	["Agents may execute tools with real-world effects","Agents may make API calls","Agents consume LLM tokens (cost)"]

AI Agents Skill

Purpose

Design and implement AI agent systems that can reason, use tools, maintain memory, and coordinate with other agents. This skill covers the ReAct pattern, tool definition, memory architectures (short-term, long-term, episodic), multi-agent orchestration, and safety guardrails. Agents are not just chatbots -- they are autonomous systems that take actions in the world.

Key Concepts

Agent Architecture

PERCEPTION:
  Input parsing, context extraction, intent classification

REASONING:
  Chain-of-thought, planning, self-reflection, error correction

ACTION:
  Tool selection, parameter construction, execution

MEMORY:
  Working memory (current conversation)
  Short-term memory (recent interactions, scratch pad)
  Long-term memory (persistent knowledge, embeddings)
  Episodic memory (past task execution records)

LOOP:
  Observe -> Think -> Act -> Observe -> Think -> Act -> ... -> Done

Agent Patterns

ReAct (Reason + Act):
  Thought: I need to find the user's order status.
  Action: query_database(order_id="ord_123")
  Observation: Order status is "shipped", tracking: "1Z999AA..."
  Thought: I have the information. I'll respond to the user.
  Answer: Your order has been shipped! Tracking: 1Z999AA...

Plan-and-Execute:
  Plan: [Step 1: Search products, Step 2: Compare prices, Step 3: Recommend]
  Execute each step, revise plan if needed

Reflection:
  After completing a task, evaluate quality and retry if insufficient

Multi-Agent:
  Researcher agent -> Analyst agent -> Writer agent -> Reviewer agent
  Each agent specialized for one part of the workflow

Patterns

Tool Definition

// Tools are typed schemas that the LLM can invoke
interface ToolDefinition {
  name: string;
  description: string;
  parameters: {
    type: 'object';
    properties: Record<string, { type: string; description: string; enum?: string[] }>;
    required: string[];
  };
}

 : [] = [
  {
    : ,
    : ,
    : {
      : ,
      : {
        : { : , :  },
        : { : , : , : [, , ] },
        : { : , :  },
      },
      : [],
    },
  },
  {
    : ,
    : ,
    : {
      : ,
      : {
        : { : , :  },
      },
      : [],
    },
  },
];