"Foreman Workflow" - Your Mini AI-Team Working On Backlog Items

What happens when you stop giving one AI assistant a task at a time—and start organizing AI agents like a development team? In this video, I walk through my experimental Foreman workflow: an orchestrator coordinates agents with separate PM, engineer, and QA responsibilities, using GitHub issues, written instructions, and a shared project board. I’m testing this on a real pet project: a platform for finding motorcycle rentals, sharing routes, and connecting riders. From my perspective as a product manager, the interesting part is how this changes the work I do. More of my attention goes into feature ideas, clear specifications, acceptance criteria, and improving the instructions the agents follow. This is a practical walkthrough of the setup—including the parts that still break. How the workflow works The Foreman coordinates the work. A PM agent refines an issue: its goal, acceptance criteria, constraints, and what belongs outside the scope. An engineer implements the task in a separate worktree. QA checks the result against the requirements and coordinates the review process before the work can move forward. The roles live in Markdown files, alongside shared process instructions and project documentation. I show how those files fit together, how agents use them, and how I’ve adjusted the rules after encountering problems. What you’ll see • Turning feature plans into smaller, tracked GitHub issues • Giving PM, engineer, and QA agents distinct responsibilities • Using worktrees to separate implementation tasks • Coordinating reviews and follow-up fixes with CodeRabbit • Tracking unrelated bugs instead of losing them in chat • Recording problems and solutions for future agent sessions • Using a GitHub project board to follow the work • Experimenting with different models and handling token limits • An unfinished model-handover plugin I’m trying to make reliable You’ll also see stalled work, inconsistent board updates, leftover worktrees, and moments when it isn’t clear whether an agent is making progress. I’m sharing an experiment, not a finished recipe for hands-off software development. Agents can handle useful work, but the quality of the plan, the instructions, and the review process still matters. The workflow itself needs attention, too. If you’re a PM, PO, or someone experimenting with AI-assisted development, I hope this gives you a concrete look at what managing an agent team can involve beyond writing a better prompt. Chapters 00:00 Why I changed my AI coding workflow 01:46 The Foreman concept: agents with defined roles 03:48 PM, engineer, and QA responsibilities 10:01 The motorcycle project behind the experiment 12:50 GitHub issues and the project board 16:25 Inside the team instruction files 22:14 Shared instructions and lessons between sessions 24:00 Model limits and the handover experiment 28:21 Starting work with the Foreman workflow 33:13 Review loops and CodeRabbit 43:40 Trying a research task live 47:29 Reviewing the groomed issue 49:48 What my job becomes 50:52 A reusable Foreman setup skill? Thank you to Alexey Grigorev and the AI Dev Tools Zoomcamp team. The course inspired me to experiment with this approach and adapt it to my own projects. Would you try this setup on a small project? Which responsibility would you delegate first: refining issues, implementing them, or reviewing the result? If you’d find a reusable skill for setting up this workflow useful, let me know in the comments. Gene Ishchuk https://ishchuk.eu #AIAgents #AICoding #ProductManagement #GitHub #ClaudeCode