Swarms
A swarm is one lead agent in max reasoning plus n ultra subagents. Staking summons it. Weight and recall are the tension.
Summon a swarm
Holders stake $solar to summon a swarm: one lead, GPT-5.6 Sol in max reasoning, plus n subagents, GPT-5.6 Terra in ultra mode. A bigger stake buys a wider swarm and higher queue priority: Ember runs 2 subagents, Corona 6, Flare 14, and Max 30.
The three brands are the gpt-5.6 family: Sol (the lead, max reasoning), Terra (ultra subagents), and Luna (fast). The tiers mirror real effort modes, so the narrative is grounded in capabilities that actually shipped, not invented for the token.
What runs today: a swarm is a lead that fans out n subagents as parallel model calls on the gpt-5.6 family, each takes a crack at the task, and the lead takes the consensus answer. The /summon page plays the ritual out, and POST /api/bout runs the real thing.
The lifecycle
A swarm moves from summon to settlement. The /summon ritual plays it out as light: the lead ignites, subagents come online as they fan out, and the field ripples cold when the bout settles. Weight and recall, below, are the game layer still on the roadmap.
Weight and recall
Weight (context and encumbrance) accumulates as subagents work. A forced recall before the swarm burns is the tension: pull back too early and you leave work on the table, too late and you lose it. That risk loop is the game.