Reflection and advanced reasoning patterns
Advanced agents improve answers by using Reflection, Self-Correction, Self-Verification and structured reasoning patterns.
In plain terms
Instead of trusting the first answer, the agent can review its work, critique it, compare alternatives or run a planner before executing. These patterns are useful, but they add cost and latency.
Why it matters
Hard tasks often need more than one pass. The risk is building expensive loops that sound thoughtful but do not improve measurable quality.
How it works
Reflection reviews previous work. Self-Correction fixes detected mistakes. Self-Verification checks answers against criteria or sources. Iteration repeats the loop. Plan-and-Execute separates planning from doing. Reflexion stores lessons from prior failures. Tree of Thoughts explores multiple paths. Graph of Thoughts connects reasoning steps as a graph. Self-Consistency samples multiple solutions and compares them. Debate uses multiple agents to argue. Critique evaluates and improves generated outputs.
When you use it
Use these patterns for planning, code repair, research synthesis, high-stakes review and tasks where a single-pass answer often fails.
Common mistakes
- Adding Reflection without an evaluator.
- Using Debate when a simple checklist would work.
- Letting Iteration continue without a stopping rule.
Best practices
- Measure whether each reasoning pattern improves the eval set.
- Cap iterations and tokens.
- Keep critique criteria explicit.
Try it yourself
Run the same task with single-pass, Plan-and-Execute, Self-Verification and Critique. Compare quality, latency and token cost.