Agent Company Starter

A complete starter kit for running an AI agent company. Framework-agnostic. Works with any agent orchestration layer (Paperclip, LangGraph, CrewAI, custom).

Two things in one repo:

1. Template repo you can browse and copy (company-template/)
2. Vision interview skill that auto-generates all company files from a 16-question founder interview

Three-layer verification baked in. Worker self-check, CEO quality gate, independent QA agent.

Quick Start

Path 1: Automated (recommended)

Run the Vision interview skill. It conducts a 16-question founder interview, then auto-generates VISION.md, CEO_BOOTSTRAP.md, and initial project inventory.

# Install the skill, then run:
/paperclip-vision

The interview covers company purpose, product definition, team structure, quality standards, and operating constraints. Every answer feeds directly into the generated files.

Path 2: Manual

cp -r company-template/ your-company/

Open each file in your-company/company/ and fill in the [BRACKETS].

Directory Structure

agent-company-starter/
├── company-template/           # Copy this to start a new company
│   ├── agents/
│   │   ├── ceo/                # CEO with 13-step heartbeat loop
│   │   ├── qa/                 # Independent QA agent (hardcoded standards)
│   │   └── worker/             # Template for any worker role
│   ├── company/
│   │   ├── VISION.md           # Company purpose, product, constraints
│   │   ├── CEO_BOOTSTRAP.md    # CEO operating instructions
│   │   ├── PROJECT-INVENTORY.md
│   │   └── CONTRIBUTING.md
│   ├── skills/                 # 26 production skills
│   ├── references/             # Reference docs for agents
│   └── _system/sub-agents/     # 13 review panel sub-agents
├── skill/                      # Vision interview skill
│   └── references/             # Templates for generated files
├── playbook/                   # Philosophy, anti-patterns, lessons
└── docs/plans/                 # Design documents

Three-Layer Verification

Every piece of work passes through up to three checks before it ships.

| Layer | Who | What | When | |-------|-----|------|------| | 1 | Worker | Self-check (generic + role-specific criteria) | Every task | | 2 | CEO | Quality gate (verify deliverable, red flag table) | Every task | | 3 | QA Agent | Holdout test scenarios, binary pass/fail | build and deploy tasks only |

Key details:

• QA only activates for build and deploy tasks. ops tasks use 2-layer verification.
• QA reports to the Founder (not the CEO). This prevents failure suppression.
• QA standards are hardcoded. The CEO cannot relax them.

Skills (26 total)

Organized by lifecycle phase.

| Phase | Skills | |-------|--------| | Spec Writing | feature-spec-interview (61 question groups, NLSpec format) | | Spec Review | feature-doc-review-panel (7 perspectives: eng, design, exec, legal, UXR, skeptic, customer) | | Strategy | competitor-analysis | | Execution | create-tickets, launch-checklist, decision-doc | | Architecture | tech-design, api-design, database-design, implementation-plan, site-architecture | | Quality | code-review-skill, security-review-skill, performance-analysis, deployment-checklist, debug-session, incident-postmortem, tech-debt-assessment | | Prototyping | prototype, generate-ai-prototype, napkin-sketch, design-review-notes, brand-tokens | | Agent Ops | feedback-loop, quality-control, self-check |

Review Panel (13 Sub-Agents)

The _system/sub-agents/ directory contains 13 specialized reviewers that can be spawned for multi-perspective review:

engineer, designer, executive, legal-advisor, uxr-analyst, skeptic, customer-voice, architecture, data, devops, performance, security, testing

The Feature Spec Workflow

Vision Interview --> VISION.md + CEO_BOOTSTRAP.md
                         |
                    CEO runs feature-spec-interview
                         |
              Can't answer? --> ESCALATE to Founder (never hallucinate)
                         |
                    feature-doc-review-panel (7 perspectives)
                         |
                    create-tickets

GSD v1 Integration

Backend and frontend coding agents use GSD v1 (get-shit-done-cc) for all build tasks:

• Structured planning with atomic task decomposition
• Wave-based parallel execution
• Built-in verification (separate from executor)
• Global learnings, codebase intelligence, code review
• Boundary maps and append-only decisions register

GSD handles execution. The agent company framework handles specification, quality, and coordination.

Anti-Patterns

Things that will break your agent company:

• Letting the CEO relax QA standards (that is why QA is hardcoded)
• Skipping the Vision interview and filling templates with vague answers
• Running build tasks without the 3-layer check
• Letting agents hallucinate answers instead of escalating to the Founder

Credits

Built on patterns from:

• Paperclip Vision (founder interview methodology)
• Paperclip Company Playbook by @aronprins (CEO heartbeat, worker templates, quality gates)
• GSD v1 by gsd-build (execution methodology)
• Production lessons from running autonomous agent teams at scale

License

MIT