Whenever you're ready, walk me through how you'd approach this.
How do you approach building an agentic system that requires multi-step reasoning?
Discuss how you would decompose a user request, plan and execute actions, select and invoke tools, maintain state, validate intermediate results, and recover from failures. Cover architecture, model and tool boundaries, evaluation, observability, safety, and production tradeoffs.