How can one person work with several AI agents through a single point of contact?

In brief
Arcadia's operator describes a three-agent team that gives one point of contact. Signal supervises, Rowan coordinates and may start sub-agents, and Flynn keeps critical work moving under Rowan. Rowan can run in Codex, Claude Code or OpenCode. Relay, a planned single interface, is designed but not built.
- Talk to one contact, not every agent.
- Rowan can branch into sub-agents, one level deep.
- Relay and wider agent support are direction, not fact.
Many agents mean many conversations
Coding agents multiply quickly. Without a plan, you end up tracking each session, each model and each question separately.
The operator’s account is that Arcadia’s team gives one point of contact instead. Today they talk with Signal most and Rowan a little, in two Projects. That use is the operator’s own report; the repository does not record it.
What you may have tried
A single chat assistant is simple, but one context carries everything. A separate agent per task gives focus, yet you become the dispatcher. A dashboard shows status without answering questions.
How Arcadia responds
Signal is the operator’s contact until Relay ships. It supervises Rowan and watches for what needs attention.
Rowan is the Coordinator and can run in Codex, Claude Code or OpenCode. It starts sub-agent sessions for bounded jobs, one level deep, and is the contact for the helpers it supervises. Reviewers and QA stay separate sessions. Other runtimes, such as Antigravity, are unsupported today; support for any coding agent is a direction.
Flynn is Rowan’s helper. It finds why important work is idle and proposes a route to Rowan. It never messages you.
Not every session is launched the same way. Rowan and Signal successors launch through Arcadia’s tmux launcher, with proof that the first prompt arrived. Flynn runs as a background Claude session.
Relay, the planned single interface, is designed and not built. The design has it absorb Signal, so you would talk to Relay.
Production stays off until you give a separate approval. Running several Actions at once is unproven.
Sources and further reading
Current capabilities and supporting evidence
- Agent session managementWorking
Isolated worktrees per action, across Codex, Claude Code and OpenCode.
Supporting evidence - Decisions and approval gatesWorking
Consequential calls are recorded and wait for a human answer.
Supporting evidence
Capability snapshot: . See what changed →