What's the difference between ReAct, Plan-and-Execute,...

Three distinct patterns, each suited to a different shape of problem:

• ReAct — reason, act, observe, repeat, deciding the next step only once the previous one’s result is known. Adaptive by construction, since every decision reacts to the latest observation, but that greediness is also its weakness: a locally reasonable choice can lead down a path that turns out to be wrong several steps later, with nothing forcing a look ahead.
• Plan-and-Execute — a planner produces the entire step sequence upfront, then an executor runs through it. This structure enables parallelizing independent steps and makes the whole run auditable before it even starts, but it’s fragile if reality departs from the plan partway through — without explicit replanning logic, the system just keeps executing a plan that no longer matches what it’s actually finding.
• Reflection — the agent generates output, critiques its own work against some quality bar, and revises, looping until it’s satisfied. This fits tasks where self-correction genuinely improves quality — writing, code generation, structured output — more than it fits multi-step action sequences, since there’s no external tool result to react to, only the agent’s own judgment of its own work.

The pattern that shows up repeatedly in production rather than picking just one: use Plan-and-Execute for the outer structure of a task whose major phases are predictable, and run ReAct inside each individual step for the adaptive execution that phase actually needs. The planner gives you auditability and parallelism at the phase level; the inner ReAct loop gives you the ability to react to a surprising tool result without the whole plan having to be redone.