The Reasoning Timing Trap in Agentic Workflows
TRIGGER
Teams implementing AI agents faced a choice between two reasoning mechanisms—extended thinking (pre-response deep reasoning) and mid-response think tools—without clear guidance on when each applies.
APPROACH
Anthropic documented the distinction: extended thinking happens before response generation and works on information available in the initial query. The think tool happens during response generation and processes information discovered via tool calls. On τ-bench, extended thinking alone achieved 0.412 pass^1 in airline domain vs 0.332 baseline—similar to unprompted think tool at 0.404. For scenarios requiring analysis of tool outputs in long chains, think tool with prompting reached 0.570.
PATTERN
“Extended thinking reasons before tool calls return; think tool reasons after. Match reasoning timing to when critical information becomes available.”
✓ WORKS WHEN
- Critical information comes from tool call results rather than the initial query (database lookups, API responses, file contents)
- Agent needs to verify compliance or correctness after each step in a chain
- You're building multi-turn agents where the model discovers constraints incrementally
- Decisions depend on combining user request with external data retrieved mid-conversation
✗ FAILS WHEN
- All information needed for the task is present in the initial user query—use extended thinking instead
- Task is primarily analytical without tool use (coding, math, physics problems)
- Agent makes single tool calls or parallel independent calls without sequential dependencies
- You need deep upfront planning before taking any actions