Role-Typed Agent Roster
Shipping a whole organisation's worth of job roles as named,
individually-invocable commands — CEO, staff engineer, design engineer,
QA lead, DX tester, chief security officer — so the operator selects a
role rather than re-describing a methodology. Each
command carries a role's working method, not just its job title: a
/qa lead doesn't only find bugs, it fixes them with atomic
commits and generates a regression test for every fix; a DX tester
actually walks the onboarding flow, times it, and screenshots the
errors; a design engineer emits shippable HTML rather than a mockup. The
roster ships as one repo and installs in one clone (Public Skill Adoption).
Two design moves make it more than a prompt collection. The first is
that the role name is the compression: "run a CEO
product review" invokes a whole stance — rethink the problem, look for
the 10-star product hiding inside the request, decide among expansion /
selective expansion / hold scope / reduction — that the operator would
otherwise have to restate every time. The role is a handle on a
methodology, which is why the roster is a Reusable Workflow Library with
a particularly memorable index. The second is splitting a role
by authority, not by method: /qa and
/qa-only run the same methodology and differ only in
whether the agent may write. That is Agent Supervision expressed as a
naming convention, and it is the roster's sharpest idea — the operator
chooses the blast radius at invocation time, in one character of
difference.
The obvious failure mode is that an org chart is not a decomposition
of your work. A roster is someone else's company, and adopting
it wholesale imports their structure along with their methods. The
/-palette after a roster install is a scrolling list, most
of which the operator will never invoke and all of which they must now
remember exists.
Claims
- A role name is a compression handle for a methodology; invoking the role is cheaper and more faithful than restating the method each time. principle — durable: this is why job titles exist in organisations too. It holds whenever the method is stable, non-obvious, and repeatedly needed.
- Split a role by authority rather than by method — same
procedure, different write permission — so the operator picks the blast
radius at invocation. best
practice — context: roles that both diagnose and remediate, where
the diagnosis is safe and the remediation is not (
/qavs/qa-only). Where a role never writes, the split is ceremony. This is Agent Supervision at the command surface. - Adopting a whole roster imports someone else's org structure, not just their methods — expect to prune it. best practice — context: installing a third-party roster built around a different company's work. The roles that fit are the ones matching work you actually do; the rest are palette clutter. See Skill Pruning and the personal-dedup discipline.
- Personas are a substrate for structured disagreement, not only for role-play. observation — the roster shares its mechanism with Adversarial Planning Council: distinct stances run in parallel over one artifact. The difference is what the diversity is for — a council's opposed personas exist to disagree, a roster's exist to divide labour. Same primitive, opposite intent.
- A roster of user-invoked commands trades context load for operator cognitive load. observation — the roster does not sit in the agent's context the way model-invoked descriptions do, but the operator must now hold a company's worth of commands in their head. The trade named in Skill Invocation Trigger, purchased in bulk.
- The canonical instance (
gstack, attributed on screen to Garry Tan, MIT-licensed) exposes roles as/-commands including/devex-review,/design-shotgun,/design-html,/qa,/qa-only,/pair-agent,/cso, and/plan-ceo-review, and its palette entries are tagged with origin and invocation mode (canary … (gstack) (user)). observation — as shown in the source; groundable, point-in-time.
Related
- Adversarial Planning Council — the same persona primitive turned to opposition rather than division of labour. A council wants its members to disagree; a roster wants them to specialise.
- Reusable Workflow Library — what a roster is, structurally: a catalog of durable workflows. The roster's contribution is that the index is an org chart, which humans navigate without a manual.
- Public Skill Adoption — how a roster travels: one repo, one clone, an entire team.
- Agent Supervision — the
/qavs/qa-onlysplit is supervision encoded as a command name. - Skill Invocation Trigger — every roster entry is an invocation decision; installing N roles is making that decision N times, by default, in the operator's favour.
- Skill Pruning — the necessary follow-up to adopting someone else's org chart.
- Layered Agentic Architecture — the roles are the "agent = scale" layer; the roster is that layer populated and named.
- Workflows vs Agents — each role is a fixed, repeatable workflow behind a name, not an open-ended agent, which is what makes it packageable.
- Execution Commoditization — a roster is a bet that the method, not the model, is what you are buying.
- Distillate: 160,000+ Cloned These 3 FREE AI Employees: Here's How (GitHub Claude Skills)