Skip to content

16. Agents

Models that pursue goals through iterative action, not single replies. Children describe the agent loop: goal → task → plan → action → observation → feedback, powered by tool use and memory, situated in an environment, optionally with a human in the loop, in single- or multi-agent configurations. Note the deliberate echo of reinforcement learning's vocabulary — an agent is RL's action/observation loop generalized to LLMs + tools.

Children

  • agent
  • goal
  • task
  • plan
  • action
  • observation
  • tool use
  • memory
  • environment
  • feedback loop
  • human-in-the-loop
  • single-agent workflow
  • multi-agent workflow

Examples

  • OpenAI Deep Research
  • Claude Code
  • Devin-like coding agents
  • OpenClaw — self-hosted personal agent runtime (LLM + tools + memory + channels + cron + real accounts)