Reasoning Techniques
≈ Algorithms / Design Patterns (e.g., Breadth-First Search, Recursion)
> Agentic Definition
Advanced cognitive architectures like Chain-of-Thought (CoT), Tree-of-Thought (ToT), and ReAct (Reason+Act) that structure the model's internal processing to solve complex problems.
> Description
Advanced cognitive architectures like Chain-of-Thought (CoT), Tree-of-Thought (ToT), and ReAct (Reason+Act) that structure the model's internal processing to solve complex problems.
≈ How It Maps to Algorithms / Design Patterns
The "method" by which the problem is solved. ToT is essentially a search algorithm (BFS/DFS) applied to the space of "thoughts."
≠ Key Divergence
Reasoning is prompted, not coded. You don't write the for loop; you tell the model how to think about the loop. You are programming the cognitive process, not the instruction set.
> Key Takeaway
Adapt: Programming is now "Prompt Engineering" at the architectural level. You are defining the algorithms of thought.
The Code
Before: Coded Algorithm
1def solve(problem):2 # Explicit implementation of A* search3 open_set = {start}4 while open_set:5 current = lowest_f_score(open_set)6 if current == goal:7 return path8 ...After: Cognitive Prompting
1prompt = """2Solve this problem using a Tree of Thought approach.31. Generate 3 possible next steps.42. Evaluate each step for feasibility.53. Discard impossible paths.64. Expand the best path.7Let's think step by step.8"""910llm.generate(prompt)Production Notes
- More reasoning steps = higher latency. CoT increases token count significantly. Use only when the task complexity demands it.
Frequently Asked Questions
When should I use the Reasoning Techniques pattern?
Advanced cognitive architectures like Chain-of-Thought (CoT), Tree-of-Thought (ToT), and ReAct (Reason+Act) that structure the model's internal processing to solve complex problems.
How does Reasoning Techniques relate to Algorithms / Design Patterns (e.g., Breadth-First Search, Recursion)?
The "method" by which the problem is solved. ToT is essentially a search algorithm (BFS/DFS) applied to the space of "thoughts." However, there is a key divergence: Reasoning is prompted, not coded. You don't write the for loop; you tell the model how to think about the loop. You are programming the cognitive process, not the instruction set.
What are the production trade-offs of Reasoning Techniques?
More reasoning steps = higher latency. CoT increases token count significantly. Use only when the task complexity demands it.