How I Run 4–8 Parallel Coding Agents

This article by Manuel Schipper details a practical system for running 4–8 parallel AI coding agents using tmux, Markdown specs, bash aliases, and six custom slash commands. The core workflow uses a role naming convention per tmux window—Planner (builds Markdown specs for features), Worker (implements from finished specs), and PM (backlog grooming)—to maintain clarity across concurrent agent sessions. Feature Designs (FDs) are structured Markdown files containing the problem, all solutions considered with pros/cons, the final solution with implementation plan, and verification steps.

Schipper identifies practical limits: around 8 agents is the maximum before losing track of decisions, and not everything parallelizes well since sequential dependencies create merge conflicts. For complex features, he uses a /fd-deep skill that launches 4 Opus agents in parallel to explore different angles before converging on a solution. Product managers benefit from understanding these parallel workflow patterns, practical scaling limits, and the role-based agent organization model applicable to managing AI-assisted development.