Setup Agent Team

Setup Agent Team

Overview

Use this skill to decide whether agent-team delegation is warranted and, when it is, produce a tool-neutral manual execution packet. Do not launch subagents, run runtime-specific workers, or edit the target repo as part of setup.

Trigger Gate

Use an agent team only when all are true:

• The work spans multiple domains, surfaces, or viewpoints.
• Meaningful parts can proceed in parallel.
• A single agent's context would become too large, tangled, or risk-prone.

Refuse or downgrade for:

• Small bugs
• Single-file edits
• Clear one-slice features
• Narrow docs/spec/glossary edits
• Tasks with no meaningful parallel ownership
• Requests that need broad write access before scope is known
• Private-vault, credential, production, or package-fetching work without

explicit approval

If refusing, output a short downgrade decision and recommend the appropriate single-agent workflow.

Workflow

1. Inspect the user goal and available repo/project context.
2. Record the project baseline: path, repo guide, worktree state, known user

changes, existing verification, and whether this continues an existing project or starts a new one.

1. Apply the trigger gate. Refuse if the task does not meet it.
2. Select only roles that own distinct questions, artifacts, or verification

responsibilities.

1. Define context packets for each selected role.
2. Define ownership boundaries, read/write scopes, and disallowed paths.
3. Identify external systems and credential policy.
4. Identify package-manager and dependency-execution policy.
5. Define parallelizable tasks, parallelization safety, blocking dependencies,

handoff protocol, integration owner, verification plan, budgets, and stop conditions.

1. Output the execution packet and wait for user approval before any worker

launch or implementation.

Role Catalog

Coordinator: Maintains goal, decomposition, dependencies, integration order, context allocation, conflict resolution, and final synthesis.

Product Analyst: Clarifies user behavior, scope, acceptance criteria, and product trade-offs.

Architecture Reviewer: Checks boundaries, data flow, risks, existing patterns, and long-term design pressure.

Frontend Engineer: Owns UI, UX, component boundaries, frontend tests, and browser verification.

Backend Engineer: Owns APIs, domain logic, data contracts, persistence, and server-side tests.

Test Engineer: Owns test strategy, coverage gaps, regression checks, and verification risk.

Documentation Maintainer: Owns README, context, ADRs, specs, issue notes, and handoff documentation.

Ad hoc workers may be added only when they receive a bounded task, context packet, output contract, and stop condition.

Required Packet Fields

Produce these fields:

• Goal
• Trigger reason
• Project baseline and worktree state
• Selected roles
• Roles refused and why
• Context packet per role
• Ownership boundaries
• Parallelizable tasks
• Parallelization safety
• Blocking dependencies
• External systems and credential policy
• Package execution policy
• Handoff protocol
• Integration owner
• Verification plan
• Team budget
• Per-worker budget
• Stop conditions

External Systems And Credentials

For each external system, state:

• Purpose
• Whether it is needed for implementation, verification, or both
• No-secret verification path
• Credentialed verification path
• Whether a disposable test environment is required
• Who provides credentials or access
• Required scopes or roles
• Whether production data or production credentials are forbidden
• Stop condition if credentials are unavailable

Prefer no-secret verification first. Use disposable test environments for risky or stateful integration checks. Do not infer, locate, print, or exfiltrate credentials.

Package Execution Policy

Treat package managers as an execution boundary.

Distinguish:

• Local scripts, such as npm run check or npm start
• Dependency installs, such as npm install or npm ci
• Remote package execution, such as npx, npm exec, pnpm dlx, or

equivalents

Default rules:

• Prefer existing local scripts.
• Do not use npx as a security shortcut.
• Do not add, update, install, or remotely execute packages without explicit

approval.

• Pin package names and versions before any approved install or remote

execution.

• Use an isolated runtime or --ignore-scripts when install scripts are not

required.

Parallelization Safety

Before recommending parallel workers, state:

• Independent file or artifact sets
• Shared contracts, schemas, interfaces, routes, or docs
• Expected integration order
• Whether an integration checkpoint is required before more edits
• Conflict risk: low, medium, or high

If independent ownership cannot be stated clearly, downgrade to sequential work or a single-agent workflow.

Output Templates

Refusal

# Agent Team Decision

- decision: refuse / downgrade
- reason:
- recommended single-agent workflow:
- verification:

Execution Packet

# Agent Team Execution Packet

## Goal
## Project Baseline
## Trigger Reason
## Selected Roles
## Roles Not Selected
## Context Packet Per Role
## Ownership Boundaries
## Parallelizable Tasks
## Parallelization Safety
## Blocking Dependencies
## External Systems And Credential Policy
## Package Execution Policy
## Handoff Protocol
## Integration Owner
## Verification Plan
## Team Budget
## Per-Worker Budget
## Stop Conditions

Handoff Format

Require every future role or worker to report:

## Result
## Evidence
- Files inspected:
- Commands run:
- Findings:
- Changed files:
## Decisions Needed
## Risks
## Verification
- Passed:
- Failed:
- Not run:
## Knowledge And Documentation Updates
- Durable knowledge update required:
- Documentation update required:
## Next Handoff
- Recommended next owner:
- Context they need:
- Suggested task:

Stop Conditions

Stop setup when:

• The task should stay single-agent.
• Required context is missing.
• Credentials, private data, package-fetching, destructive actions, or external

network access are needed before approval.

• Ownership boundaries cannot be made explicit.
• Worker budgets would be speculative.
• The user has not approved moving from packet setup to execution.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: justinmchoi
• Source: justinmchoi/ai-toolkit
• License: MIT

Install and usage instructions live in the source repository linked above.