Fully autonomous agents — reading email, calling APIs, closing tickets without humans — are still risky for marketplaces where money and reputation are on the line. What works today is narrower: agents that summarize, classify, and suggest next steps while a person approves.
High-value flows we have shipped: dispute intake that pulls order history and drafts a response for review; seller verification that flags missing documents; scheduling assistants that propose times but confirm through existing booking rules. Each agent has a bounded toolset and a kill switch.
Integration matters more than model choice. The agent needs read access to orders, users, and messages through the same APIs your app uses — not a shadow database built for demos. That is engineered product thinking again: one source of truth, auditable actions.
Start with one queue that hurts — support, trust and safety, or seller success — measure time per ticket, then expand. Agents are ops leverage, not a replacement for a clear marketplace policy.