Multi-step reasoning: guardrail → plan → retrieval → subagents → answer. The activity timeline and todo checklist surface inline as the agent works.