Skip to content

Agent Dispatch

Management has specialist sub-agents for focused work. Dispatch with:

management: agent <slug> <task>

Available specialists

Slug Use for
product-manager PRD shaping, scope trade-offs, feature prioritization rationale
product-sprint-prioritizer Slicing scope into sprints, sequencing within a roadmap phase
project-management-project-shepherd Cross-department coordination, dependency unblocking, status rollups

Agent persona files live in:

  • .claude/agents/<slug>.md
  • .codex/agents/<slug>.md
  • .opencode/agents/<slug>.md

The three files for a given slug must be identical. When you edit one, sync the other two. (See Handoff for why tool parity matters.)

When to dispatch vs do it yourself

Dispatch when:

  • The task is narrow and benefits from a focused persona (e.g. "prioritize these 12 user stories into 3 sprints").
  • You want an opinionated second pass on an existing draft.
  • The main session is already deep in context and the sub-task is independent.

Do not dispatch when:

  • The task is a routine doc generation already covered by a prefix (prd:, srs:, etc.).
  • You haven't run the interview yet — interview first, dispatch later if needed.

What the agent must do

Every dispatched agent runs with caveman mode on and writes its output to the path declared in dept.yaml → outputs. The agent does not invent new output locations.