Agent Architecture — Overview

Our agent system uses a hierarchical architecture with a single Orchestrator (the "super agent") that holds member and trip context, and a set of Specialized Research Agents that handle domain-specific work delegated to them.

The Orchestrator does not perform research itself. It decomposes user requests, dispatches scoped tasks to the appropriate specialists, and synthesizes their outputs into a coherent response.

Architecture

Orchestrator — Owns all persistent state. Decides what needs to happen, who does it, and how the results combine.

Research Agents — Stateless specialists. Each owns a focused domain (e.g., dining, hotels, transport). They receive a narrow brief, do the work, and return a structured result.

Chapters

1. Orchestrator — the super agent
2. Orchestrator Tools — the internal cascade
3. Job Queue (Redis/Valkey) — the message bus
4. Research Agents — abstracted specialists
5. Job Lifecycle — request to response
6. Scaling — horizontal and sharded
7. Failure Modes — what breaks and how we recover
8. Deployment — container layout
9. Observability — traces, metrics, logs