Skip to content

AI Agents Overview

AI agents are autonomous systems built on large language models (LLMs: neural networks trained on massive text corpora to generate and reason over language). They perceive, plan, and act toward a goal. Unlike a chatbot that answers one prompt at a time, an agent runs a perceive–plan–act loop: observe the environment, decide, act, and repeat until the task is done. As Anthropic's engineering team puts it: "Agents... are systems where LLMs dynamically direct their own processes and tool usage, maintaining control over how they accomplish tasks."

What Makes an Agent Different from a Chatbot?

Aspect Chatbot Agent
Interaction Single prompt → response Multi-step workflow with tool use
State Stateless or session-only Persistent state across sessions
Autonomy Follows explicit instructions Makes decisions, handles ambiguity
Tools None or limited File I/O, shell, APIs, code execution

Core Agent Capabilities

Most AI agents (what Lilian Weng calls LLM-powered autonomous agents) include:

  1. Planning – Breaking complex tasks into subtasks
  2. Tool Use – Calling external APIs, reading/writing files, running commands
  3. Memory – Short-term (context window) and long-term (external storage)
  4. Reflection – Self-correction based on feedback or errors
  5. Orchestration – Coordinating multiple sub-agents for parallel work

When to Use Agents

Reach for an agent when the task involves multi-step reasoning (research, coding, debugging), tool integration (file work, API calls, DB queries), iteration over build → test → fix cycles, or long-running work that spans sessions. For simple Q&A or one-shot content generation, a plain LLM call is usually enough. For an applied walkthrough of an agent loop in a developer tool, see Claude Code: an agentic CLI in practice.

Single-Agent vs. Multi-Agent

Architecture Best For Trade-offs
Single Agent Focused tasks, simpler orchestration Limited parallelism
Multi-Agent Complex workflows, specialized roles Higher coordination overhead

Multi-agent helps when you need distinct roles (Planner, Coder, Reviewer) or can parallelize the work. Anthropic's Building effective agents (2024) recommends starting with a single agent and only moving to multi-agent when the trade-offs clearly justify the coordination cost. For a working example on Azure, see OpenClaw — a multi-agent reference implementation on Azure.