Agent Planning

Definition

Agent planning is the capability of an LLM agent to decompose complex goals into executable sub-tasks, reason about dependencies and sequencing, and dynamically adjust strategies based on intermediate outcomes.

智能体规划是 LLM 智能体将复杂目标分解为可执行子任务、推理依赖关系和顺序、并根据中间结果动态调整策略的能力。

Details

Planning transforms agents from reactive responders into proactive problem-solvers. It involves three core capabilities:

1. Task Decomposition Breaking down high-level goals into smaller, manageable steps. Example: “Book a flight to Paris” → [Search flights, Compare prices, Select seat, Enter payment, Confirm booking]

任务分解：将高层目标分解为更小、可管理的步骤。例如："预订去巴黎的航班" → [搜索航班、比较价格、选择座位、输入付款、确认预订]。

2. Dependency Reasoning Understanding which sub-tasks must complete before others can begin. Some tasks can run in parallel; others have strict ordering constraints.

依赖推理：理解哪些子任务必须在其他任务开始前完成。有些任务可以并行运行；其他任务有严格的顺序约束。

3. Dynamic Replanning Adjusting the plan when sub-tasks fail or new information emerges. Requires Reflection to assess what went wrong and generate alternative strategies.

动态重新规划：当子任务失败或出现新信息时调整计划。需要反思来评估出了什么问题并生成替代策略。

Planning Methodologies

Hierarchical Decomposition Top-down approach where each goal is recursively broken into sub-goals until reaching atomic actions.

分层分解：自顶向下方法，每个目标递归分解为子目标，直到达到原子操作。

Reasoning Frameworks

• Chain-of-Thought (CoT): Step-by-step reasoning before action
• Tree of Thoughts (ToT): Exploring multiple reasoning paths in parallel
• ReAct (Reason + Act): Interleaving reasoning and action execution

推理框架：思维链（CoT，行动前逐步推理）、思维树（ToT，并行探索多条推理路径）、ReAct（推理+行动，交错推理和行动执行）。

Evaluation Benchmarks

Current research evaluates planning capabilities using:

• TaskBench: Focuses on task decomposition, tool invocation, parameter prediction
• AgentBench: Multi-environment evaluation of reasoning and decision-making
• Planet: Workflow automation and multi-step execution benchmarks
• PARTNR: Planning in embodied multi-agent household scenarios

评估基准：当前研究使用 TaskBench（关注任务分解、工具调用、参数预测）、AgentBench（多环境推理和决策评估）、Planet（工作流自动化和多步执行基准）、PARTNR（具身多智能体家庭场景规划）。

Connections

• Related to: Reflection (replanning requires reflection on failures)
• Related to: Skills (planning determines which skills to invoke and when)
• Related to: Cognitive Depth Adaptation (planning is System 2 thinking)
• Mentioned in: Deep Agents

---

Sources added by Heal on 2026-04-06:

• ArXiv - TaskBench · 2026-04
• GitHub - Awesome-LLM-Planning · 2026-04
• Microsoft - Agent Planning Research · 2026-04