Databases

• Postgres — tickets, agents, work items, and the per-call token and dollar table. Row-level security scopes user-owned data to its owner; app-owned domain state is keyed to the user who owns it.
• Redis — the internal mesh: agent streams, consumer groups, heartbeats and runtime coordination.
• Vector store — the retrieval corpus, embedded server-side. If a collection is ever missing its embedding function, retrieval degrades to lexical scoring rather than silently returning nothing, which is the failure that is actually dangerous.
• Graph tier — engine-agnostic, with a per-person and a per-tenant database. For relationship-heavy domains it beats a pile of joins.

It is absent unless you configure it, and the graph API answers a plain 503 when it is not there rather than pretending. Relational data stays relational. A new domain graph is ingestion plus queries over the same connector — never a new service — and forcing graph where relational fits is explicitly the wrong call.

Per-call tokens and dollars land in the cost table as the work happens, attributed to the bot that spent them. When you ask what an agent cost you last month, the answer is a query against the system of record — not an estimate reconstructed from logs.