Run a solo company with eight agent departments

Tools

GitHub Copilot

Overview

Posted on Mar 3

Workflow

A solo founder in Portugal built 8 AI agent departments (CEO, CFO, COO, Lawyer, Accountant, Marketing, CTO, Improver) using GitHub Copilot custom agents to manage 5 SaaS products. Agents share a persistent knowledge graph via MCP memory servers and consult each other automatically based on trigger tables. The COO runs daily standups to delegate tasks, while the Improver agent analyzes lessons to upgrade other agents. The human founder reviews proposed changes as diffs before merging, ensuring no autonomous modification of agent instructions.

Reported result

After two months, the system produced 84+ tweets and 5 articles, reducing marketing time to under 1 hour/week and ensuring zero missed compliance deadlines. However, revenue remained minimal at €6.09. The author notes that chaining 10 steps with 85% accuracy per agent drops overall accuracy to 20%, and recommends starting with only 3 agents (COO, Marketing, Accountant) to cover 80% of value. Memory corruption issues were resolved by forking the MCP server to add async mutexes and atomic writes.

Human role

The original post describes the human role this way: Now it proposes changes as diffs that I review before merging — it can't modify other agents autonomously.

Limitations

Prerequisites include GitHub Copilot, MCP servers for tools (X, Sentry, etc.), and a management repo with .agent.md files. Constraints include context window limits, hallucination risks requiring inter-agent review, and memory corruption from parallel writes. The system is not a replacement for strategic thinking; humans must decide on pivots and approve all changes. Proposed advice: Start with 3 agents, invest in memory structure early, and test agent outputs against each other to catch hallucinations.

Try it

I am a solo founder managing multiple SaaS products. I want to set up a multi-agent system using GitHub Copilot custom agents. 1. Create a management repo with a .github/agents/ directory. 2. Define 3 initial agents: COO, Marketing, and Accountant. Each should be a .agent.md file with specific domain instructions (e.g., Marketing knows my voice and schedule; Accountant knows local tax deadlines). 3. Set up an MCP memory server to store a knowledge graph (entities: product, decision, lesson; relations: owns, depends-on). 4. Configure inter-agent communication: Marketing must call Lawyer for claim reviews; CFO must call Accountant for tax checks. Implement a max depth of 3 to prevent loops. 5. Create a COO agent that runs a daily standup: checks Sentry errors, scans sprint board, reads knowledge graph, and delegates tasks. 6. Add an Improver agent that reads 'lesson' entities from memory to suggest updates to other agents' instructions. 7. CRITICAL SAFEGUARD: The Improver must only propose changes as diffs. I will manually review and merge these diffs. No agent can autonomously modify another agent's instructions. First, perform a dry run: Have the COO agent generate a daily plan based on a sample Sentry error log and sprint board state. Show me the plan and the proposed task delegations. Do not execute any real actions (posting, filing, spending) until I confirm the output is correct. If inputs are missing (e.g., specific tax rules), ask me for them before proceeding.

More cases · Office & operations

• Answer dental clinic calls and manage appointments
• Email scheduled project summaries with Codex
• WorkBuddy: Automated Daily Work Hour Logging
• Clean inbox and draft social posts in your voice
• Handle routine email tasks and follow-up reminders with a personal assistant