Agentic AI 4 min

Beyond ReAct: Planning Patterns

ReAct isn't the only way to structure an agent. Compare it against Plan-and-Execute and Reflexion, and know when each is worth the extra complexity.

ReAct decides one step at a time: think, act, observe, repeat. That is not the only way to structure an agent. Two other patterns show up constantly in production systems: Plan-and-Execute and Reflexion.

Plan-and-Execute
Generate the full plan upfront (a numbered list of steps), then execute each step, only replanning if a step fails or reveals new information. Fewer expensive reasoning calls to the large model, because planning happens once instead of before every single action. The tradeoff: the plan can be wrong about steps 4 and 5 before you have any evidence from steps 1 through 3.
Reflexion
Run the task, then have the agent critique its own output against the goal, and retry with that critique added to context if it falls short. This adds a self-correction loop on top of any other pattern. The tradeoff: at least one extra full pass (and its cost) for every attempt, in exchange for meaningfully higher success rates on tasks where the first attempt is often subtly wrong.
Planning a Road Trip

ReAct is deciding your next turn only when you reach each intersection. Plan-and-Execute is mapping the entire route before you leave, and only pulling up maps again if a road is closed. Reflexion is arriving, checking whether you actually reached the right city, and turning back if you did not.

When to use which. ReAct: exploratory tasks where each observation genuinely changes what you should do next, like research or debugging. Plan-and-Execute: tasks with a knowable structure, like 'refactor these 5 files' or 'set up this deployment,' where thinking between every micro-step wastes calls. Reflexion: tasks where correctness matters more than latency or cost, like a code review agent or anything shipping to a user unsupervised. Nothing stops you from combining them: plan upfront, execute each step with a ReAct loop, reflect at the end.

What's Next
Every one of these patterns needs somewhere to put what the agent has learned so far. That's memory, and it is the single most common cost mistake in agent systems.