Multi-agent systems love role cards: planner, critic, researcher, coder. Some of that specialization is real engineering; some is theater with personas. The difference is whether the role changes what the agent can do — tools, context, and standards — or merely how it introduces itself.

Specialization pays through three mechanisms, none of which require a new model. Context focus: a code-reviewer agent whose window contains only the diff, the relevant subsystem, and the review checklist outperforms a generalist assistant holding an entire repo and a chat history — focus raises signal density. Tool authority: a "deployer" role that is the only agent holding deploy credentials turns organizational policy into architecture. Evaluable standards: a role with a written definition of done can be tested — "did the reviewer actually flag the four seeded defects in the eval set?" — where a vibe-based persona cannot.

The evidence on persona-only prompting is more mixed, which is the honest caveat: elaborate personality cards often move output style more than output quality, and multi-agent setups where every agent has the same tools and the same context frequently underperform one good agent with a checklist. The gains consistently show up when roles partition work, not adjectives.

Specializing for Real

• Partition the context, not just the title: each role's window should contain what its job needs and exclude what it does not.
• Bind authority to roles: capabilities follow the role — the fence from your workspace architecture is drawn around roles, not individuals.
• Write the standard, then test it: a role exists when you can state its acceptance criteria and measure an agent against them.
• Prefer few roles: every boundary adds a handoff cost; add a role only when a measured failure — attention dilution, authority leakage — demands it.

The organizational echo is exact: team design teaches that specialization creates leverage when it narrows interfaces, and creates bureaucracy when it merely multiplies titles. Agent fleets, built by engineers who sat through those retrospectives, should know better by default.