Multi-step reasoning
The ability of an LLM or agent to chain multiple inferences together to solve a problem — answer A leads to question B, which leads to question C, and so on until the final answer.
Multi-step reasoning is what separates real problem-solving from pattern-matching. A model that can answer "What was Einstein's favorite food?" might fail at "Of the foods that Einstein's favorite chef was famous for, which originated in Italy?" — the second requires chaining facts across steps.
In 2026 the dominant approach is chain-of-thought prompting combined with reasoning models like o3 and Claude Sonnet 4.6 with extended thinking. These models internally generate intermediate reasoning steps before answering, dramatically improving accuracy on multi-hop questions.
For agents, multi-step reasoning is foundational. Every agentic loop — plan, act, observe, act again — is multi-step reasoning in practice. The quality of the reasoning at each step determines whether the agent ships useful work or wanders.
Frequently asked
How do I improve multi-step reasoning in my agent?+
Use a reasoning model (o3, Claude with extended thinking) for the planning step. Force explicit chain-of-thought in the prompt. Decompose the problem into sub-questions. Use [tree of thoughts](/glossary/tree-of-thoughts) when the right path is hard to find in one shot.
Why do LLMs fail on multi-step problems?+
Cumulative error. If each step is 95% accurate, a 5-step chain is only 77% accurate (0.95^5). Reasoning models offset this with verification at each step; non-reasoning models compound errors.