WritingEngineering
Twenty-seven specialist roles
Why we route AI work to specialist roles instead of asking one generalist to do everything.
Ask an assistant to “make this faster” and the obvious move is to start editing code. But that request is really a measurement job first. Agent Dispatcher is our answer to that mismatch: it reads the work being asked for, routes it to the best-fit specialist role and works the way that role works.
Route by the work, not the topic
There are 27 roles, including UI/UX designer, reviewer, tester, debugger, architect, researcher, product manager, security auditor, data analyst and copywriter. The dispatcher matches the request against what each role is for, and also against what it is not for. When two roles look close, the deciding question is the deliverable: a decision, a diff, a verdict, a reproduction and fix, a measured speedup, a test suite or a screen.
Chains, with limits
Real requests often hold more than one kind of work, so roles can chain inside a turn: a planner hands to an implementer, an explorer hands to a debugger. Each role meets its own definition of done before handing over. Three roles per turn is the practical ceiling, and the chain stops at the deliverable that was actually asked for.
A verifier never carries the role that did the work
When work fans out to several agents, each one is told its role, its slice of the job and what is out of scope. The role that produced the work is never the one that judges it: an implementer writes, and a reviewer or tester checks. Several copies of the same role pointed at the same evidence give you one blind spot several times over, so we give verifiers different lenses on purpose.
Where you’ll find it
Locus 3.3 ships 27 specialist agent templates, and Agent Dispatcher brings the relevant specialist guidance into Work, Plan and Grill chats. You can start from a template or build a custom role, and edit its instructions, model, mode and permissions.
Draft — for reviewOctober 2026 · Engineering