A reasoning pattern is the shape of an agent's thinking: the order in which it plans, acts, observes and checks its own work. The simplest is ReAct, where the model alternates between a thought and a tool call. Others plan the whole task first and then execute it, write and run code instead of calling one tool at a time, or critique and revise an answer before returning it.
No pattern is best for every task. A short lookup needs one tool call; a multi-step research job benefits from an explicit plan; data work often goes faster as code. These guides start with an overview of the main patterns and when to use each, then build them one by one: ReAct from scratch, plan-and-execute, code-action agents, self-improving agents that learn from past runs, and governed self-modification.
The code in these guides runs on Promptise Foundry, an open-source Python framework for AI agents and MCP servers.