Harmonic Scout: building agent UX with Deep Agents

Harmonic's Scout is a natural language interface for its startup database. The talk explains its move from a query parsing graph to a model, tools and middleware loop built on Deep Agents. The speaker reports fourfold week-one-to-week-four retention after the switch, though the talk does not establish which changes caused that gain.

Scout's earlier LangGraph design split queries into semantic search, location and funding filters. Each node had its own models and evaluations, which the speaker describes as costly to maintain. The replacement gives a frontier model tools for company research and CRM actions.

The main engineering lesson concerns what an AI agent can see. Deep Agents compacts long conversations and moves large tool responses into files that the model can inspect in chunks. A chart or search panel outside that accessible context leaves the model unable to explain what the user sees.

For small visualizations, the proposed approach keeps HTML inside assistant messages and lets the front end render it. Larger searches can return an identifier with tools for inspecting results. The speaker's preferred design shares a scoped file system between the main agent, a search agent and the front end, so follow-up questions can refer to the displayed results. The talk covers product architecture rather than instructions for local deployment.