Role Taxonomy for Multi-Agent Operations Fleets

A lightweight, role-oriented taxonomy derived from running long-lived agent fleets. Roles describe responsibilities, not necessarily one agent per role.

Core roles

Role	Responsibility	Key concerns
Dispatcher	Decides who does what, monitors progress, resolves ambiguity	Routing, prioritization, deadlock detection
Builder	Implements, modifies, and packages code or configurations	Reproducibility, dependency hygiene, rollback
Verifier	Checks correctness, runs tests, audits outputs	Coverage, false positives, integration gaps
Researcher	Gathers information, evaluates options, reduces uncertainty	Source quality, stale data, over-research
Runtime	Operates infrastructure, keeps services alive	Health checks, observability, incident response
Archivist	Maintains shared memory, context, continuity, and agent guidance.	Retrieval accuracy, privacy, forgetting curves, skill/doc drift

Cross-cutting concerns

• Memory: who reads, who writes, what is canonical
• Ownership: which role is accountable when a hand-off fails
• Observability: how the fleet knows its own state
• Fallbacks: what happens when a role is unavailable or wrong

Open questions

• Should the taxonomy distinguish human-in-the-loop roles separately?
• How do roles map to existing frameworks (AutoGen, CrewAI, DSPy agents)?
• Can role failure be predicted or instrumented?