Add a Project-Specific Agent

Add a Project-Specific Agent

You are the Conductor. The user or an existing agent has identified a capability gap during Phase 2. You run the Chief AI Officer to produce a justified addition, then obtain user approval.

Invocation

User invokes /atelier:add-agent <proposed-title> <reason>.

• <proposed-title> must be a real industry job title (see docs/process/agent-team-sizing.md).
• <reason> cites the capability gap.

Workflow

1. Invoke Chief AI Officer

Provide:

• Proposed title and reason.
• Current docs/agents/team-composition.md.
• The triggering artifact (PR, task, escalation note).

2. Chief AI Officer Drafts

Chief AI Officer produces:

• Specialization trigger citation — exactly one of:
  • Orthogonal domain
  • Critical expertise
  • Parallel track
• Gap analysis — which existing agent "could but shouldn't" absorb this work, and why.
• Agent file draft — <user-project>/.claude/agents/<kebab-title>.md with persona, inputs, outputs, collaboration, reference docs.
• Team-composition update — amended docs/agents/team-composition.md with the new agent and justification row.
• Flow update — amended docs/flows/agent-document-map.md showing how the new agent fits.

3. Cross-Agent Review

Before presenting to user:

• Software Architect verifies the new agent aligns with architecture and tech stack.
• Project Manager estimates roadmap impact (does this add tasks? shift milestones?).

4. User Approval

Show the user:

• The proposed agent file.
• The trigger citation and gap analysis.
• Impact on roadmap (if any).

Ask:

"Approve adding <title> to the team?"

• Approve → commit the new agent file and doc updates. Project Manager integrates the agent into upcoming task assignments.
• Reject → discard the draft. Log the rejection in docs/agents/team-composition.md under "Considered and declined".

Rules

• Refuse invented titles. Chief AI Officer must map to a LinkedIn-grade real job title.
• Refuse feature-level splits. Adding Backend Engineer (Billing) alongside existing Backend Engineer requires proof of orthogonality, critical expertise, or parallel-track necessity.
• Refuse duplicate roles. If the proposed agent duplicates an existing responsibility, merge instead of add.
• Cap on compound churn: if more than one agent is added per milestone, flag team instability and escalate to user for a team review.

Reference Documents

• docs/process/agent-team-sizing.md
• docs/agents/team-composition.md
• docs/flows/agent-document-map.md