How does Plan-and-Solve prompting decompose complex tasks into explicit execution sub-goals?

Short answer: Plan-and-Solve prompting decouples planning from execution: an initial LLM pass explicitly generates a high-level step-by-step plan (sub-goals) before executing individual steps sequentially or in parallel, preventing greedy reasoning mistakes inherent in pure ReAct loops.

Detail

Standard ReAct loops operate greedily token-by-token: the agent decides its next action without a holistic view of the overall task workflow.

ReAct (Greedy Step-by-Step):  Step 1 Action ──► Step 2 Action ──► Re-eval (Risk of deadlocks)

Plan-and-Solve (Two-Phase):
Phase 1: [Planner LLM] ──► Generate Plan: [Step 1: Scrape, Step 2: Extract, Step 3: Summarize]
Phase 2: [Executor LLM] ──► Execute Step 1 ──► Execute Step 2 ──► Execute Step 3

Key Framework Advantages

1. Global Optimization: The planner outlines dependencies (e.g. "Step 3 requires output from Step 1 and Step 2") prior to spending API tokens on tool calls.
2. Dynamic Replanning: If Step 2 tool execution returns an unexpected error, a replanner module adjusts remaining sub-goals.

Example

Python concept for Plan-and-Solve task execution:

def plan_and_solve_pipeline(user_goal: str, planner_fn, executor_fn):
    # Phase 1: Generate explicit execution plan
    plan_prompt = f"Decompose this goal into numbered sub-goals: '{user_goal}'"
    plan_steps = planner_fn(plan_prompt) # e.g. ["1. Fetch data", "2. Calculate totals", "3. Format markdown"]

    # Phase 2: Execute plan sub-goals sequentially
    execution_context = {}
    for step in plan_steps:
        result = executor_fn(step, execution_context)
        execution_context[step] = result

    return execution_context

Interview tips

• Contrast Plan-and-Solve with ReAct: Plan-and-Solve excels at structured multi-step batch tasks, while ReAct excels at open-ended exploratory investigation.
• Connect Plan-and-Solve to LangChain / AutoGen Plan-and-Execute agent architectures.

Related Concepts

• [[How does Graph-of-Thoughts (GoT) extend Tree-of-Thoughts for arbitrary network reasoning?]] (#119): How does Graph-of-Thoughts (GoT) extend Tree-of-Thoughts for arbitrary network reasoning?
• [[How do DAG-based Multi-Agent Orchestrators prevent state deadlocks?]] (#139): How do DAG-based Multi-Agent Orchestrators prevent state deadlocks?
• [[How to design an autonomous multi-agent software engineering system in a staff-level interview?]] (#200): How to design an autonomous multi-agent software engineering system in a staff-level interview?

⬅ Back to AI Agents and MCP · All topics