CEO Agent Workflow

A file-driven agent orchestration framework that enables technical startups to run their operations through Claude Code.

Overview

CEO Agent Workflow provides a structured approach to AI-powered business operations. Instead of ad-hoc prompting, you define your company's state, goals, and constraints in markdown files—then let a CEO agent delegate work to department agents (engineering, product, ops, growth) through a systematic workflow.

This is not about replacing human judgment. It's about amplifying human capability by automating the mechanical translation of business decisions into coordinated action across your organization.

Why Use This?

Traditional approaches to AI assistance treat language models like search engines—you describe a goal, get a response, and hope it fits your context. CEO Agent Workflow inverts this: your company state becomes the source of truth that drives execution.

The problem with ad-hoc AI assistance:

• Context is lost between sessions
• No continuity across business functions
• Human review happens too late (or not at all)
• Decisions aren't logged or traceable

What CEO Agent Workflow provides:

• Persistent company state that agents read and respect
• Coordinated delegation across specialized department agents
• Built-in human review gates for sensitive actions
• Decision logging and audit trails
• File-driven architecture—everything is inspectable markdown

Core Philosophy

CEO Agent Workflow shares philosophical roots with GitHub's spec-kit, but applied to business operations rather than software specifications:

1. Files Over Prompts: Company state, agent definitions, and decisions live in version-controlled markdown files—not ephemeral chat histories.

2. Human-in-the-Loop: The CEO agent flags sensitive actions (payroll changes, public releases, expenses over thresholds) for human approval before execution.

3. Specialization Through Agents: Rather than one general-purpose assistant, specialized department agents handle their domains with focused context and capabilities.

4. Transparency Over Magic: Every decision is logged. Every agent's instructions are readable. Nothing happens in a black box.

Quick Start (End Users)

Prerequisites

• Claude Code installed and authenticated

Note: The framework uses an Obsidian-style vault structure (organized markdown files). The Obsidian app is optional but recommended for browsing and editing your company knowledge base.

Installation

# Clone this repository
git clone https://github.com/alessio-locatelli/ceo-agent-workflow
cd ceo-agent-workflow

# Initialize a new project (creates in current directory)
./src/init_startup.sh my-startup

# Or specify a custom path
./src/init_startup.sh ~/projects/my-startup
./src/init_startup.sh /path/to/my-startup

This creates a directory with the complete framework structure at your specified location.

First Run

1. Configure your startup by editing:

   • COMPANY_STATE.md — your mission, goals, team, and budget
   • .claude/CLAUDE.md — repository URLs and connector configs

2. Run the CEO agent:

   /company/run-ceo
   

3. Review outputs in:

   • CEO/daily-run-{{DATE}}.md — session summary
   • inbox/FROM_HUMANS.md — items flagged for your review

See the generated project's README.md for complete usage documentation.

Upgrading an Existing Project

If you bootstrapped your project previously and want to get the latest framework features (updated agent prompts, new slash commands, etc.) while preserving your company data:

# From the bootstrap repository
./src/upgrade.sh /path/to/your-existing-project

The upgrade script will:

• ✅ Update all agent definitions and slash commands
• ✅ Update framework templates and configuration files
• ✅ Preserve your company data (COMPANY_STATE.md, connectors config)
• ✅ Preserve all agent-generated content (logs, specs, meetings, etc.)
• ✅ Require a clean git state (commit your changes first)

Files that will be updated:

• .claude/agents/*.md (fresh agent prompts)
• .claude/skills/, .claude/scripts/, .claude/rules/git.md
• vault/templates/, vault/.obsidian/, vault/README.md
• Root README.md

Files that will be preserved:

• .claude/rules/connectors.md (your connector configuration)
• .claude/rules/project-location.md (your project location)
• vault/COMPANY_STATE.md (your company data)
• All agent-generated content in vault/ (inbox/, meetings/, specs/, etc.)
• Your customized .gitignore, .pre-commit-config.yaml, etc.

After upgrading, review changes with git diff and commit if everything looks good.

Current Status

⚠️ Pre-alpha: This project is in early development. The API, file structure, and agent definitions may change significantly. We welcome feedback and contributions, but recommend waiting for a stable release before production use.

Currently, only Claude Code is supported.

Want support for another AI coding agent (Codex, Gemini CLI, Open Code)?

We’re happy to support more services and tools! Please open an issue or submit a pull request.

Default Agents (can be changed or extended by a user)

Agent	Purpose
CEO	Daily operations, strategic decisions, delegation
Engineering	Code implementation, PRs, bug fixes
Product	Specs, priorities, customer feedback
Ops	Infrastructure, deployments, cost management
Growth	Analytics, conversion optimization (post-MVP)

Human Review Gates

The CEO agent automatically flags these actions for human approval:

• Payroll or billing changes
• Public releases
• File or folder deletion
• External API state changes
• Expenses exceeding $500
• Mission or focus changes
• Hiring new agents