Agents, teams & voice
// recruit specialists, run them concurrently, stay in command
Recruiting
New agents come through the Recruitment Bay. Each recruit has a dossier — class, loadout, personality, and the skills it ships with. A short interview at creation shapes what the agent knows about you and your goals; the dossier keeps growing as you work together.
Classes and archetypes
StarNet ships a roster of built-in classes (researcher, builder, marketer, analyst, and more), each with a loadout tuned to its job. None are gated — every archetype is available from the start, and every agent draws on the station's shared gear.
The dossier learns you
The dossier isn't static config — it's a living study of its commander. Agents notice patterns in how you work and what you approve, and propose beliefs ("prefers a working demo over a written spec"); each proposal is yours to accept or strike, and the full ledger of what your crew believes about you is readable and editable at any time. Ratings you give on finished work feed a taste profile, so delivery style converges on what you actually value. Nothing is learned behind your back: proposals are explicit, and struck beliefs stay struck.
The model is a roster property
Each agent is pinned to a model on the roster — not per-chat. Put your heavy reasoning model on the architect and a fast cheap model on the intern, and route work accordingly. Swap an agent's model at any time in its dossier.
Concurrency
Each additional bay or specialist you place can run concurrently — genuinely distinct, bounded agent runs, not one queue wearing costumes. The conveyor carries real work items between rooms.
Delegation: agents that summon agents
An agent holding the task delegation capability can break a large job apart and hand the pieces out. Every worker it creates runs its own real agent loop — its own searching, reading, file writing and memory — and returns a result the lead folds back into the answer. There are three shapes:
| Move | What it does |
|---|---|
| Dispatch | Hand subtasks to specialists already on your crew. |
| Spawn | Create ephemeral clones of itself to work subtasks in parallel, then discard them. |
| Summon | Recruit a genuinely new specialist onto the crew — the same thing you'd do in the Recruitment Bay, done live because the job needed it. |
Background workers are inspectable while they run: list them, look at one, interrupt one that's going the wrong way, or resume a stalled one — it restarts on the same prompt and appends to the same durable record, so nothing is silently lost. You watch the whole tree work on the station floor; a lit desk is a live worker, not a decoration.
Live Voice — talking to the station
You can put COMMS down and just talk. Hands-free Live Voice listens, transcribes, runs the work, and speaks the answer back, using a speech engine built into the app — Whisper for listening, Kokoro for speaking, both on your CPU.
It deliberately does not use a vendor's realtime voice API, even where one is available. A vendor path gave smoother turns but tied the feature to one provider, capped the persona behind that vendor's guardrails, and made the station behave differently depending on which provider you happened to connect. The built-in engine is a little slower per turn and wins everything else: no key, no account, identical on every station, and your recorded audio never leaves the machine. The first session may download the open model weights once. After that, listening is offline; speaking is offline while Kokoro is available and otherwise may use the keyless Microsoft Edge Read Aloud fallback for answer text.
COMMS
The COMMS window is command: talk to any agent, stream its work live, approve consequential actions, and answer the (rare, bounded) clarifying question before a big task commits. When an agent needs a decision it asks once with concrete choices — it can't ask and quietly mutate things at the same time; the harness blocks write tools until the question is settled.