R02 / Run
Specialist roster
Meet Guild Stack's roster: 15 domain specialist templates minted on demand, plus the advisor and developer machinery agents, for scoped lanes and typed handoffs.Page evidence and guidance
Use this page to understand who Guild Stack can assign to a run, why only a few specialists are selected at a time, and where you can approve or change the proposed team before work starts.
Try it: /guild:plan proposes a phase team, writes the team artifact, and waits at the team approval gate before dispatch.
- product framing
- scope review
- approved spec
- frontend lane
- backend lane
- typed handoffs
- adversarial review
- verification
- quality verdict
Who ships
Since the v2.2 roster change (shipped on the beta next channel), Guild Stack registers 2 machinery agents with the host and ships 15 domain specialist type templates.
| Group | Roles | What they are for |
|---|---|---|
| Engineering | architect, researcher, backend, frontend, mobile, devops, qa, security | Design, code, tests, infrastructure, release readiness, and risk review. |
| Content and communication | copywriter, doc-writer, technical-writer, social-media, seo | Public copy, product docs, doc-site pages, search, and communication work. |
| Commercial | marketing, sales | Positioning, campaigns, outbound, proposals, and discovery material. |
| Machinery (registered) | advisor, developer | Escalated judgment and generic implementation lanes. |
The 15 domain specialists are the durable roster, shipped as read-only type templates at templates/specialists/<role>.md in every host package rather than as host-registered agents. The two machinery agents are the only agents Guild Stack registers with the host. They are augmenting: dispatchable, but excluded from domain matching and the team-size count. advisor answers narrow escalations at the powerful tier, and developer covers implementation work with no better domain owner.
Each specialist has a trigger boundary and a skill family, carried by its template and inherited verbatim by minted instances. Every role definition also carries persona frontmatter: an operating_style plus personality axes for terseness, pushback posture, and escalation bias. A backend lane should not quietly become deployment work; a copy lane should not become API reference; a security lane challenges risk instead of implementing the fix. Completed lanes return structured handoff receipts, so the next specialist can see what was done and what remains.
How a team is proposed
Guild does not put every role on every team. It composes a small team per phase by matching spec domains against your project’s specialist instances in .guild/agents/ plus the shipped template library. A domain specialist joins a team as a project instance: a template-covered domain with no instance yet is minted deterministically into .guild/agents/<role>.md — a byte-preserving copy of the template, stamped derived_from_template, with no approval gate because templates are pre-blessed. An existing instance is always reused, never re-created. Dispatch does not need a host-registered agent under the domain name; lanes run through a host-generic subagent type that adopts the minted definition.
The team artifact is:
.guild/team/<slug>.<phase>.yaml
The composer matches the phase, spec domains, and existing specialists. If it finds a genuinely novel domain the template library does not cover, it asks at the team gate whether to create, skip, substitute, or compose from scratch. Selecting create supplies explicit human authority: a genuinely new role does not need prior runs or reflections, but it still must pass its interview, boundary scan, applicable boundary evals, and new-role paired evals. Learning can also propose candidates; those evolution-proposed roles retain the historical extraction and shadow requirements and are never silently promoted into the live roster.
The usual team is 3-4 specialists. The default cap is 6; the machinery agents (advisor, developer) do not count toward it. Larger teams are possible, but Guild Stack prints the cap warning because bigger teams create more coordination and review work.
Beginner examples
If you ask for a new API endpoint, Guild Stack might propose:
| Phase | Likely team | Why |
|---|---|---|
| Planning | architect, backend, qa | Decide the shape, implementation lane, and test evidence before code starts. |
| Development | backend, qa, maybe security | Build the endpoint, test it, and challenge auth or external-integration risk if present. |
| Quality | qa, devops, maybe security | Check release readiness instead of continuing feature implementation. |
If you ask for website copy, Guild Stack might propose copywriter, seo, and doc-writer. If the request includes a public launch, marketing may join. If the work turns into implementation, frontend owns the UI lane instead of the copy specialist.
The important part: the team changes with the phase. The planning team can be right, and still not be the right quality team later.
Power-user controls
Power Users can inspect and steer the proposal before it becomes the run contract:
- Team file:
.guild/team/<slug>.<phase>.yaml - Plan lanes:
.guild/plan/<slug>.md - Approval gate: the G-team gate during planning
- Size control:
/guild:plan --team-size=N - Role gap options: create, skip, substitute, or compose from scratch
- Tier posture: each lane has a routing tier; the exact model id resolves through
.guild/settings.json - Capability scope: lanes may carry a tool allow-list; absent means no added lane-specific scope
Team composition is not the dispatch backend. The backend mode is resolved at run start, while the team file records who owns each lane for the current phase.
Roster reference
The first fifteen rows are the domain specialist templates; advisor and developer are the registered machinery agents.
| Role | Typical trigger | Boundary |
|---|---|---|
architect | System design, ADRs, tradeoffs | Does not implement the feature. |
researcher | Research, comparison, papers, vendors | Does not make the architecture call. |
backend | APIs, data, queues, workers, migrations | Does not own frontend, mobile, or infra release work. |
frontend | Web UI, client state, accessibility, bundling | Does not own backend contracts or mobile clients. |
mobile | iOS, Android, React Native, Expo | Does not own backend APIs. |
devops | CI, deploy, observability, incidents, runbooks | Does not own application logic. |
qa | Test strategy, regression, flaky tests, snapshots | Does not write the feature itself. |
security | Threats, auth, secrets, CVEs, audits | Challenges and validates risk; fixes route to implementers. |
copywriter | Product copy, microcopy, voice, email | Does not own API docs or technical reference. |
doc-writer | Product docs, doc-site pages, README-style guides | Does not own API reference or marketing copy. |
technical-writer | API docs, manuals, tutorials, release notes | Does not own marketing copy. |
social-media | Posts, threads, calendars, engagement templates | Does not own long-form docs or campaigns. |
seo | Keywords, metadata, internal links, crawl health | Does not own the core copy decision. |
marketing | Positioning, launches, campaign briefs, A/B variants | Does not own sales outreach. |
sales | Cold outreach, discovery, proposals, follow-up | Does not own launch strategy. |
advisor | One narrow escalated question | Sees the draft and question, not raw context. |
developer | Generic implementation lane | Defers to domain implementers when one fits. |
Related
- Architecture & Lifecycle, where the machinery agents and specialist templates sit in the lifecycle, and the skill taxonomy each role loads.
- Agent Communications, the handoff envelope and receipt contracts in detail.
- Cost-Aware Model Tiering, how each lane’s tier is auto-scored, plus advisor escalation.
- Self-Evolving Skills & Specialists, the minting lifecycle for project-specific specialists.