AI Prompt Engineering and ReACT Framework | LinkedIn
Claude.ai take
Chain of Thought (CoT) vs ReAct Frameworks
Chain of Thought (CoT)
Mechanism: The model generates intermediate reasoning steps before producing a final answer. The reasoning occurs entirely within a single inference pass.
Structure:
Input → Reasoning Step 1 → Reasoning Step 2 → ... → Final Answer
Key characteristics:
- Pure text generation, no external tool calls
- All reasoning happens in the model's forward pass
- Can be elicited via prompting ("Let's think step by step") or fine-tuning
- Variants include zero-shot CoT, few-shot CoT, and self-consistency (sampling multiple chains and aggregating)
Limitations:
- Cannot access external information
- Cannot verify intermediate conclusions
- Reasoning quality degrades on problems requiring factual lookup or calculation
ReAct (Reasoning + Acting)
Mechanism: Interleaves reasoning traces with action execution in a loop. The model reasons about what to do, executes an action (tool call), observes the result, then reasons again.
Structure:
Thought → Action → Observation → Thought → Action → Observation → ... → Final Answer
Key characteristics:
- Explicitly separates reasoning (Thought) from tool use (Action)
- Observations ground the reasoning in external reality
- Loop continues until the model determines it has sufficient information
- Original paper (Yao et al., 2022) demonstrated improvements on knowledge-intensive tasks
Implementation pattern:
Thought: I need to find X
Action: search("X")
Observation: [search results]
Thought: Based on these results, Y is relevant
Action: lookup("Y")
Observation: [lookup results]
Thought: I now have enough information
Answer: ...
Current Relevance (End of 2025)
What's changed:
Native tool use — Modern models (Claude, GPT-4+, Gemini) have tool calling built into the API layer rather than requiring prompt-based ReAct loops. The model outputs structured tool calls directly; the orchestration layer executes them. This replaces the text-based "Action:" parsing.
Extended thinking / reasoning tokens — Models now support explicit reasoning phases (Claude's extended thinking, OpenAI's o1/o3 reasoning tokens) that are architecturally distinct from output tokens. This supersedes prompt-based CoT for complex reasoning.
Agent frameworks abstract the pattern — LangGraph, CrewAI, AutoGen, and similar frameworks implement ReAct-style loops as infrastructure. You configure tools and let the framework handle the thought-action-observation cycle.
Multi-step planning has evolved — Approaches like tree-of-thought, graph-of-thought, and Monte Carlo Tree Search over reasoning steps have emerged for problems requiring exploration rather than linear chains.
Where they still matter:
- CoT remains relevant for single-inference problems where you want interpretable reasoning without tool overhead
- ReAct's conceptual model (reason → act → observe → reason) is still the dominant pattern for agents, just implemented differently
- Understanding these frameworks helps you debug agent behavior and design tool schemas effectively
What's arguably obsolete:
- Text-based ReAct parsing ("Action:", "Observation:" markers) — replaced by structured tool calls
- Manual CoT prompting for capable models — extended thinking handles this natively
- Simple linear ReAct loops — complex tasks now use hierarchical planning, reflection, and retry mechanisms
