How it works: The agent decomposes a goal into steps, selects the right tool for each, executes them in sequence, and synthesizes a final result — while always escalating sensitive items to a human. Plan → tool calls → reasoning trace → action → human handoff.