Specialist agents

The roster

Role	Pulls tickets typed	What it does	Model surface
chief_of_staff	planner tickets	Decompose briefs, dispatch, replan	Frontier-only
ba_agent	ba tickets	Research, user interviews, requirement refinement	Frontier or local
architect	architect tickets	Technical design spikes, tradeoff memos	Frontier
dev_agent	dev tickets	Implementation (code-gen via AgentHub daemon)	Local CLI subscription
qa_agent	qa tickets	Test plans, manual QA checklists, bug triage	Local or frontier
memory_optimizer	memory tickets	Summarise long histories, manage context budgets	Frontier

Each role has:

• A system prompt (tenant-editable via Settings → Roles).
• A skills list (vendored; which skills this role can reach).
• A subagents list (specialists this role can delegate to).
• A budget cap (monthlyBudgetTokens on the AgentRole row).

ba_agent

User research, discovery, requirements.

• Reads the brief, the meeting transcript, and any attached documents.
• Produces a structured output: user segments, pain points, open research questions.
• Can spawn a sub-ticket for follow-up user interviews (marked as needing human involvement).

Prompt tuning: terse, no hedging. Research is about finding specifics, not summarising generalities.

architect

Design spikes and tradeoff memos.

• Reads the brief plus the project graph’s god-nodes.
• Produces: candidate approaches, tradeoff table, recommended direction, risks.
• Typical output: 3 approaches compared on 4 axes, a recommendation paragraph, a list of “what we don’t know yet.”

Prompt tuning: explicit about uncertainty. “We don’t know yet” is a valid answer; “recommended approach” requires justification.

dev_agent

Code synthesis via the AgentHub daemon on your local machine. This is where Workforce0 differs from most AI products: we DON’T call Anthropic’s API for code-gen. We call your Claude Code / Cursor / local CLI on your laptop.

Flow:

1. dev_agent ticket enqueued.
2. Agent daemon on your machine picks it up via WebSocket.
3. Daemon runs claude (or equivalent) with the ticket’s prompt + your local repo context.
4. Output (diff, test results, notes) POSTs back to the backend.
5. Ticket transitions to done or failed.

See Agent daemon for the full flow.

Why not the API?

• Subscription vs per-token. Your Claude Code subscription is paid. Double-paying via the API is dumb.
• Source-code locality. Production code never leaves your machine. No transit to a third party beyond the CLI’s own call.
• Existing trust stack. You already trust Claude Code with your repo. Extending that trust to Workforce0 is additive.

When no agent daemon is running

Tickets stay in the dev queue and re-dispatch when an agent comes online. A warning banner in the UI alerts the installer.

qa_agent

Acceptance tests, manual QA, bug triage.

• Reads the brief plus the dev_agent’s output.
• Produces: a test plan, an acceptance criteria checklist, notes on untested edge cases.
• Can spawn sub-tickets for specific bug fixes.

Often runs on a local model (Ollama / vLLM) — QA output is less quality-sensitive than planner output.

memory_optimizer

Context-budget management. Summarises long ticket histories, rolls up comment threads, keeps the planner’s context under the model window limit.

Rarely runs on demand — usually triggered by a size threshold on the ticket’s accumulated state.

Adding a new role

Contributors, see Adding an agent role. The general recipe:

1. Add a row to the agent_roles table (via seed or UI).
2. Register a queue name in backend/src/queues/.
3. Optionally write a BaseConsultant subclass in backend/src/services/agents/. (Purely-LLM-driven roles don’t need a subclass.)
4. Add a Prisma migration if the role needs tenant-scoped config.
5. Ship.

Routing tickets to roles

The chief-of-staff writes roleSlug on each step. The backend enqueues into the matching queue. A ticket with an unknown role stays pending forever (and logs a warning).

Role vs subagent

• Roles pull tickets. They’re queue consumers.
• Subagents are specialists loaded into a role’s context. A dev_agent ticket might include the code-reviewer subagent as extra context for the code-gen prompt.

See Skills & subagents.