Back to full roadmap
topiccore
Reasoning Models in Agents
o1/o3, Claude thinking, Gemini 2.5 — revolutionary for agentic planning but 5-10× latency/cost.
3 hours2 resources1 prereqs
Reasoning models shine in agents at two points:
- Planning — decomposing complex tasks
- Tool selection — picking the right tool among plausible candidates
Pattern: "Reasoning model as planner, fast model as executor". E.g. Claude Opus thinking builds the plan (1 call, $0.05), GPT-4o-mini executes each subtask (10 × $0.001).
Trade-off: reasoning takes 10-60s per step. A 5-step loop → 1-5 min total. For UX: streaming + progress indicator + "First plan ready" milestone.
Anthropic extended thinking budget: thinking_tokens: 8000 lets the model think up to 8K tokens. include_thoughts: true exposes the reasoning for debugging.