Skip to main content
Curia is a team of agents, not a single assistant. The design is deliberate: a single Coordinator handles every conversation while specialist agents do the focused work behind the scenes. The person on the other end never knows multiple agents were involved. Diagram showing the multi-agent architecture: You connects to the Coordinator inside the Bullpen, which delegates to Calendar, Contacts, CEO Inbox, and Research specialists. All agents draw from an Execution Layer of 65+ skills.

​
The three tiers

You → Coordinator. The Coordinator is the only agent that communicates outward — with you, with your contacts, with anyone. It holds your configured persona (display name, tone, email signature) and presents a single, coherent face for everything Curia does. There is exactly one Coordinator per deployment. The Bullpen. All agents — the Coordinator and every specialist — operate inside the Bullpen, a structured internal channel where agents can delegate, hand off, and coordinate. When the Coordinator needs research done or a calendar conflict resolved, it dispatches a task to the relevant specialist inside the Bullpen and waits for the result. Every exchange is logged with full causal tracing: you can always see who asked whom to do what and why. The Execution Layer. Agents don’t have direct access to external systems. They request actions — send an email, look up a contact, search the web — through the Execution Layer, which validates permissions against each skill’s manifest before running anything. This hard boundary means a misbehaving agent can’t do anything its declared skills don’t allow.

​
Why this structure

The coordinator-specialist pattern solves two problems at once. Coherence. Having one agent own every outbound message means the persona stays consistent even when six different specialists contributed to the answer. The Coordinator synthesizes, re-voices, and delivers — the underlying specialization is invisible to the recipient. Isolation. Each specialist has its own memory scopes and permissions. The research analyst can’t read the CEO inbox; the calendar specialist can’t write to the contacts database. A compromised or misbehaving specialist can only affect the narrow scope it was given.

​
Extensibility

Specialists are just YAML files. Adding a new one — a contracts analyst, a travel planner, a finance tracker — is a matter of writing a manifest and dropping it in the agents/ directory. The add more slot in the diagram is literal: there’s no architectural limit on the number of specialists, and the Coordinator can discover new ones dynamically if allow_discovery: true is set.

How agents work

Deep dive into agent YAML configuration, TypeScript handlers, error budgets, and skill discovery.

The Bullpen

How agents coordinate internally — threaded, logged, and interruptible from any channel.