Team orchestration

Team Builder

Interactive menu for browsing and composing agent teams on demand. Works with flat or domain-subdirectory agent collections.

When to Use

• You have multiple agent personas (markdown files) and want to pick which ones to use for a task
• You want to compose an ad-hoc team from different domains (e.g., Security + SEO + Architecture)
• You want to browse what agents are available before deciding

Prerequisites

Agent files must be markdown files containing a persona prompt (identity, rules, workflow, deliverables). The first # Heading is used as the agent name and the first paragraph as the description.

Both flat and subdirectory layouts are supported:

Subdirectory layout — domain is inferred from the folder name:

agents/
├── engineering/
│   ├── security-engineer.md
│   └── software-architect.md
├── marketing/
│   └── seo-specialist.md
└── sales/
    └── discovery-coach.md

Flat layout — domain inferred from shared filename prefixes. A prefix counts as a domain when 2+ files share it. Files with unique prefixes go to "General". Note: the algorithm splits at the first -, so multi-word domains (e.g., product-management) should use the subdirectory layout instead:

agents/
├── engineering-security-engineer.md
├── engineering-software-architect.md
├── marketing-seo-specialist.md
├── marketing-content-strategist.md
├── sales-discovery-coach.md
└── sales-outbound-strategist.md

Configuration

Agents are discovered via two methods, merged and deduplicated by agent name:

1. claude agents command (primary) — run claude agents to get all agents known to the CLI, including user agents, plugin agents (e.g. SKC:architect), and built-in agents. This automatically covers SKC marketplace installs without any path configuration.
2. File glob (fallback, for reading agent content) — agent markdown files are read from:
   • ./agents/**/*.md + ./agents/*.md — project-local agents
   • ~/.claude/agents/**/*.md + ~/.claude/agents/*.md — global user agents

Earlier sources take precedence when names collide: user agents > plugin agents > built-in agents. A custom path can be used instead if the user specifies one.

How It Works

Step 1: Discover Available Agents

Run claude agents to get the full agent list. Parse each line:

• Plugin agents are prefixed with plugin-name: (e.g., SKC:security-reviewer). Use the part after : as the agent name and the plugin name as the domain.
• User agents have no prefix. Read the corresponding markdown file from ~/.claude/agents/ or ./agents/ to extract the name and description.
• Built-in agents (e.g., Explore, Plan) are skipped unless the user explicitly asks to include them.

For user agents loaded from markdown files:

• Subdirectory layout: extract the domain from the parent folder name
• Flat layout: collect all filename prefixes (text before the first -). A prefix qualifies as a domain only if it appears in 2 or more filenames (e.g., engineering-security-engineer.md and engineering-software-architect.md both start with engineering → Engineering domain). Files with unique prefixes (e.g., code-reviewer.md, tdd-guide.md) are grouped under "General"
• Extract the agent name from the first # Heading. If no heading is found, derive the name from the filename (strip .md, replace hyphens with spaces, title-case)
• Extract a one-line summary from the first paragraph after the heading

If no agents are found after running claude agents and probing file locations, inform the user: "No agents found. Run claude agents to verify your setup." Then stop.

Step 2: Present Domain Menu

Available agent domains:
1. Engineering — Software Architect, Security Engineer
2. Marketing — SEO Specialist
3. Sales — Discovery Coach, Outbound Strategist

Pick domains or name specific agents (e.g., "1,3" or "security + seo"):

• Skip domains with zero agents (empty directories)
• Show agent count per domain

Step 3: Handle Selection

Accept flexible input:

• Numbers: "1,3" selects all agents from Engineering and Sales
• Names: "security + seo" fuzzy-matches against discovered agents
• "all from engineering" selects every agent in that domain

If more than 5 agents are selected, list them alphabetically and ask the user to narrow down: "You selected N agents (max 5). Pick which to keep, or say 'first 5' to use the first five alphabetically."

Confirm selection:

Selected: Security Engineer + SEO Specialist
What should they work on? (describe the task):

Step 4: Spawn Agents in Parallel

1. Read each selected agent's markdown file
2. Prompt for the task description if not already provided
3. Spawn all agents in parallel using the Agent tool:
   • subagent_type: "general-purpose"
   • prompt: "{agent file content}\n\nTask: {task description}"
   • Each agent runs independently — no inter-agent communication needed
4. If an agent fails (error, timeout, or empty output), note the failure inline (e.g., "Security Engineer: failed — [reason]") and continue with results from agents that succeeded

Step 5: Synthesize Results

Collect all outputs and present a unified report:

• Results grouped by agent
• Synthesis section highlighting:
  • Agreements across agents
  • Conflicts or tensions between recommendations
  • Recommended next steps

If only 1 agent was selected, skip synthesis and present the output directly.

Rules

• Dynamic discovery only. Never hardcode agent lists. New files in the directory auto-appear in the menu.
• Max 5 agents per team. More than 5 produces diminishing returns and excessive token usage. Enforce at selection time.
• Parallel dispatch. All agents run simultaneously — use the Agent tool's parallel invocation pattern.
• Parallel Agent calls, not TeamCreate. This skill uses parallel Agent tool calls for independent work. TeamCreate (a Claude Code tool for multi-agent dialogue) is only needed when agents must debate or respond to each other.

Examples

User: team builder

Claude:
Available agent domains:
1. Engineering (2) — Software Architect, Security Engineer
2. Marketing (1) — SEO Specialist
3. Sales (4) — Discovery Coach, Outbound Strategist, Proposal Strategist, Sales Engineer
4. Support (1) — Executive Summary

Pick domains or name specific agents:

User: security + seo

Claude:
Selected: Security Engineer + SEO Specialist
What should they work on?

User: Review my Next.js e-commerce site before launch

[Both agents spawn in parallel, each applying their specialty to the codebase]

Claude:
## Security Engineer Findings
- [findings...]

## SEO Specialist Findings
- [findings...]

## Synthesis
Both agents agree on: [...]
Tension: Security recommends CSP that blocks inline styles, SEO needs inline schema markup. Resolution: [...]
Next steps: [...]

---

Team Agent Orchestration

Use this skill when agents are being managed like a team rather than a single assistant. The purpose is to make team-based orchestration reliable: clear work items, explicit ownership, agent Kanban state, branch isolation, control pane visibility, and merge gates.

When to use

• The task spans multiple agents, tools, harnesses, branches, or worktrees.
• The user mentions team orchestration, agent Kanban, squad, conductor, control pane, manager, desktop app, Zellij, tmux, Hermes, Devin, Codex, Claude Code, or multi-agent work.
• A project needs shared workflow state across people and agents.
• Existing agent fan-out is producing output but not mergeable product.

Operating Model

Treat every agent as a teammate with a narrow contract:

• Owner: the person or agent accountable for the work item.
• Scope: files, branch, tool surface, and forbidden areas.
• State: backlog, ready, running, review, blocked, merged, or archived.
• Evidence: tests, screenshots, logs, review notes, or eval reports.
• Merge gate: the exact condition that allows integration.

Agent Kanban

Use agent Kanban when work must be visible across sessions.

Column	Meaning	Exit Criteria
Backlog	Candidate work item, not yet shaped	Acceptance criteria written
Ready	Shaped and assignable	Owner and branch/worktree assigned
Running	Agent is actively working	Handoff artifact and changed files exist
Review	Work is complete but not merged	Tests, diff review, and risk check pass
Blocked	Needs external input or failed gate	Blocker has owner and next action
Merged	Integrated into mainline	PR merged or local main updated
Archived	No longer relevant	Reason recorded

Each card should fit this schema:

{
  "id": "agent-card-001",
  "title": "Build dynamic workflow skill",
  "owner": "codex",
  "state": "running",
  "branch": "product/dynamic-workflow-team-orchestration",
  "worktree": ".",
  "acceptance": [
    "Skill exists",
    "Tests cover required concepts",
    "Content artifact contains video and article angles"
  ],
  "merge_gate": "lint, focused tests, and catalog check pass",
  "handoff": "path/to/handoff.md"
}

Team-Based Orchestration Flow

1. Shape the board: convert fuzzy ambition into work items with owners and merge gates.
2. Pick execution mode: single-agent, dynamic workflow mode, MPX/tmux, worktree fan-out, or external desktop orchestrator.
3. Assign boundaries: one owner per card, clear file scope, and no overlapping writes without an integrator.
4. Run agents: each agent writes evidence and handoff notes, not just code.
5. Review in sequence: tests first, then diff review, then security/risk checks, then content/product polish.
6. Merge deliberately: one integrator resolves conflicts and updates the control pane or status artifact.
7. Extract reusable skill: if the card pattern repeats, promote it into skills/.

Control Pane Requirements

A useful control pane for team orchestration should show:

• Active work items and their agent Kanban state.
• Owner, harness, branch, worktree, and last heartbeat.
• Links to handoff artifacts, tests, screenshots, and PRs.
• Blockers grouped by owner and unblock action.
• Merge readiness by gate, not vibes.
• Reusable workflow candidates that should become shared skills.

Do not add more automation until the operator can answer: who owns this, what changed, what gate failed, and what can safely merge?

Dynamic Workflow Compatibility

When a card needs dynamic workflow mode:

• Put the task-local framework under the card owner.
• Store inputs and outputs on the card.
• Require an eval before moving from Running to Review.
• Promote the harness to a shared skill only after repeat use.

Failure Modes To Watch

• Agent soup: many agents running, no owner or merge gate.
• Invisible work: useful output exists only in a chat transcript.
• Board theater: a Kanban board exists but cards have no acceptance criteria.
• Overlapping writes: parallel agents edit the same files without worktrees.
• No product artifact: the process produces docs but no runnable or publishable surface.

Output Standard

Finish each orchestration pass with:

• Board/card changes.
• Merged or pending branches.
• Tests and eval evidence.
• Blockers with owner and next action.
• New shared skill candidates.