ReAct, chain of thought and reflection in practice

Most of what frameworks call "agents" is a handful of patterns with new names. Chain-of-thought writes intermediate steps before the answer. Self-consistency samples several reasoning paths and votes. ReAct interleaves thoughts with tool actions and real observations. Plan-and-execute plans first and runs steps with cheaper calls. Tree of Thoughts searches over reasoning branches. Reflexion and self-critique generate, evaluate and retry. Each buys something (intermediate work, diversity, contact with reality, structure, search, feedback) with tokens, calls and latency. Reasoning models have absorbed explicit step-by-step thinking, but acting through tools, planning across calls and checking against external signals remain your design choices. This file owns the patterns as ways the model reasons; the program around it (state, termination, step budgets) is the agent loop in The agent loop, written by hand.

• Every pattern is a way of spending tokens. Know the multiplier before you adopt it.
• ReAct's value is the observation, not the thought. Reasoning checked against real tool results beats reasoning in a vacuum.
• Critique needs an external signal. Tests, validators and evidence make reflection work; "are you sure?" often does not.
• The model proposes; deterministic code executes. Tools, arguments, observations and budgets are owned by your code.
• With reasoning models, drop explicit CoT; keep acting, planning and checking.