Agent Prompts

Disclaimer. Tool-use APIs, parallel tool calling, and computer-use APIs vary by provider and evolve quickly. The prompt patterns here are stable; the wire format is the provider's documentation.

A chatbot answers a question. An agent takes actions to achieve a goal — calls tools, observes results, decides what to do next. The agent loop is Plan → Execute → Observe → Adapt, and the prompts that drive it are the difference between a useful agent and a model that flails for 30 turns and gives up.

The composition model

Prompts are atoms.
Skills are molecules.
Agents are structures.

An agent is built from layers of prompts: