Plan-and-Execute
Plan-and-Execute is an agent task planning pattern in which a model first generates a complete execution plan and then carries it out step by step, dynamically adjusting based on feedback received during execution.
Core Facts
Most mainstream coding agent architectures adopt the ReAct (Reasoning + Acting) paradigm.
AI Agents typically adopt the ReAct (Reasoning + Acting) pattern, iterating in a loop between reasoning and acting.
Agent planning capabilities are typically implemented through the ReAct (Reasoning + Acting) framework.
Timeline (last 90 days)
ReAct, Plan-and-Execute, and Reflexion are three common agent control-flow patterns, differing in who decides the next step; they are essentially the harness's orchestration choices for loop structure, termination conditions, and feedback loops, not model capabilities.
In the Plan-and-Execute pattern, the model first produces a complete plan in one go, and then the harness executes it step by step; controllability is strong, but when plan quality is poor, the whole thing fails.
ReAct, Plan-and-Execute, and Tree of Thoughts are three different agent planning and task decomposition patterns, whose core difference lies in when and at what granularity planning occurs.
About 90% of Plan-and-Execute's token consumption is in the executor loop rather than in the planning text.
Plan-and-Execute consumes more than 2.3 times the tokens of ReAct and is dominated across the board by the other three designs.
ReAct consumed an average of 4,663 tokens per question, Plan-execute 10,535 tokens, Supervisor 7,494 tokens, and Draft-verify 4,905 tokens.
In the 240-question test, ReAct's accuracy was 59.6%, Plan-execute's 61.7%, Supervisor's 65.0%, and Draft-verify's 63.7%.
Codex's 'plan mode' corresponds to the Plan-and-Execute framework in AI Agent design, planning the execution checklist before executing, and belongs to the Human-in-the-loop mechanism.
Agent systems often adopt a combination of "pre-retrieval in the planning phase + on-demand retrieval in the execution phase"; the former ensures the overall direction is correct, while the latter supplements context in specific sub-steps.
Agent-driven architectures are commonly referred to as ReAct or Plan-and-Execute patterns, and have matured rapidly in recent years as LLM capabilities have improved.
12 more timeline events
All Facts (20)
Most mainstream coding agent architectures adopt the ReAct (Reasoning + Acting) paradigm.
90%VerifiedAI Agents typically adopt the ReAct (Reasoning + Acting) pattern, iterating in a loop between reasoning and acting.
90%VerifiedAgent planning capabilities are typically implemented through the ReAct (Reasoning + Acting) framework.
90%VerifiedPlan-and-Execute 框架先由规划器生成完整的多步任务计划，再逐步交由执行器完成，适合任务边界清晰的场景
80%VerifiedPlan-and-Execute和ReAct是当前Agent领域两种代表性的规划范式
80%VerifiedThe demo project uses a primary agent (Hub) to coordinate three to four sub-agents running four blocks in parallel, embodying a Plan-and-Execute plus sub-agent division-of-labor architecture.
75%VerifiedReAct is more flexible and can dynamically adjust strategies, while Plan-and-Execute is more stable in long task chains, reducing global deviation caused by local errors.
75%VerifiedModern agents typically operate based on the ReAct (Reasoning + Acting) paradigm, where the model first reasons about the current state and then decides which tool to call.
75%VerifiedThe ReAct pattern's execution flow includes six steps: the user raises a request, the agent reasons (Reasoning), executes an action (Acting), observes the result (Observation), loops repeatedly, and returns an answer
70%VerifiedMainstream task planning methods include the ReAct paradigm, Plan-and-Execute mode, and Tree of Thoughts
65%VerifiedThe Plan-and-Execute pattern first has a planner generate a complete execution plan, which is then carried out step by step by an executor; it is more stable in long tasks but places extremely high demands on the quality of the planning phase
65%VerifiedThe Plan-and-Execute paradigm is suitable for long-horizon complex tasks, while ReAct is more efficient in scenarios with high tool-calling frequency that require real-time feedback.
65%VerifiedThe Plan-and-Execute paradigm first generates a complete plan and then executes it step by step, making it suitable for long-horizon tasks.
65%UnverifiedThe ReAct paradigm has spawned variants such as Reflexion, LATS, and Plan-and-Execute.
60%UnverifiedThe Plan-and-Execute architecture decouples planning from execution, reducing loop risk, but lacks the ability to dynamically correct course.
60%UnverifiedMainstream Agent frameworks such as ReAct and Plan-and-Execute can both use MCP as a standardized implementation of the tool-calling layer.
60%UnverifiedIn the Plan-and-Execute pattern, the model first produces a complete plan in one go, and then the harness executes it step by step; controllability is strong, but when plan quality is poor, the whole thing fails.
50%UnverifiedReAct, Plan-and-Execute, and Reflexion are three common agent control-flow patterns, differing in who decides the next step; they are essentially the harness's orchestration choices for loop structure, termination conditions, and feedback loops, not model capabilities.
50%UnverifiedReAct, Plan-and-Execute, and Tree of Thoughts are three different agent planning and task decomposition patterns, whose core difference lies in when and at what granularity planning occurs.
50%UnverifiedAbout 90% of Plan-and-Execute's token consumption is in the executor loop rather than in the planning text.
50%