Multi-Turn Reasoning: Agents That Think Across Steps

Beyond one-shot prompting

A single prompt forces the AI to generate an answer in one go. For hard problems — proofs, long code, research — even the best model gets some parts wrong on the first pass. Multi-turn flows let the model reflect, critique itself, and retry.

The generate-critique-revise loop

TURN 1 (generate): 'Write a 200-word argument for why schools should start later. Make it persuasive.' --> draft v1 TURN 2 (critique): 'You are a skeptical peer reviewer. List the three weakest points in the argument above. Be specific.' --> critique list TURN 3 (revise): 'Using the critique, revise the argument. Address each of the three weaknesses. Keep it under 200 words.' --> draft v2A classic three-turn improvement cycle.

Each turn plays a different role — author, critic, editor. Studies on Claude and GPT consistently show that a two- or three-pass structure outperforms a single pass, especially on reasoning-heavy tasks. The model is effectively its own reviewer.

The ReAct pattern (reason + act)

ReAct interleaves reasoning steps with tool calls. Instead of one big plan, the model thinks a little, uses a tool (search, calculator, code execution), reads the result, thinks again, and so on. This is the foundation of modern AI agents.

THOUGHT: I need to find the CEO of Notion and their hiring trends. ACTION: search("Notion CEO") OBSERVATION: Ivan Zhao is CEO of Notion Labs. THOUGHT: Now I need recent hiring data for Notion. ACTION: search("Notion hiring 2026") OBSERVATION: Notion announced 200 new engineering roles in Q1 2026. THOUGHT: I have enough to answer. FINAL: Notion, led by CEO Ivan Zhao, announced 200 new engineering roles in Q1 2026.ReAct loop. Thought / Action / Observation repeats until Final.

Reflection on tool failures

1. If a tool errors, don't just retry blindly — have the model reason about why and adjust inputs.
2. If a tool returns empty results, try a broader query before giving up.
3. If results contradict each other, surface the contradiction rather than silently picking one.
4. Cap the loop — set a max of N iterations so the agent can't spin forever.

Keeping context manageable

Multi-turn flows accumulate tokens quickly. Strategies: summarize older turns into a running memo; store long data in a scratchpad tool; use XML tags to section the conversation. A 1M-context Claude can hold a lot, but cost and latency still grow.