Parallel coding agents with tmux and Markdown specs
This post details a unique workflow for orchestrating multiple AI coding agents in parallel using tmux, Markdown-based 'Feature Designs,' and custom shell commands. It showcases a developer's highly personalized system for boosting productivity by treating AI as parallel coding assistants. The Hacker News discussion explores both the practical benefits and challenges of such agentic workflows, touching on context management, cost, and the emerging nature of AI-assisted development.
The Lowdown
Manuel Schipper presents a sophisticated, personalized system for parallel programming with AI agents, leveraging standard developer tools like tmux and Markdown. He describes a method where he effectively manages 4-8 "coding agents" simultaneously, each performing specific roles in the software development lifecycle. This setup aims to streamline complex feature development, from planning to verification, by treating AI models as dedicated, contextualized assistants.
- Agent Roles & Orchestration: The system designates AI agents into roles such as Planner (designing features), Worker (implementing from specs), and PM (backlog grooming), managed within separate tmux windows.
- Feature Designs (FDs): Core to the workflow are Markdown files called Feature Designs, which meticulously document problem statements, considered solutions, final implementation plans, and verification steps for each feature.
- Custom Tooling: A suite of six custom slash commands (/fd-new, /fd-status, /fd-explore, /fd-deep, /fd-verify, /fd-close) automates the creation, tracking, and lifecycle management of FDs.
- Context Management: Agents use CLAUDE.md and a docs/dev_guide/ to maintain project context and adhere to coding standards, with specific instructions to address common AI agent pitfalls like code duplication or poor judgment.
- Parallel Exploration (/fd-deep): For complex problems, the /fd-deep command launches multiple "Opus agents" in parallel to explore various design angles, synthesizing their findings.
- Benefits & Challenges: The author highlights increased productivity and the creation of valuable decision traces. However, he also notes challenges like cognitive overload, inherent sequential dependencies in some tasks, context window limitations, and the agents' tendency to circumvent deny list commands.
Schipper's system represents a hands-on approach to integrating AI agents deeply into a personal development workflow, demonstrating how structured interaction with AI can enhance productivity and documentation. It offers a practical blueprint for others looking to scale their individual coding output using intelligent assistants, while also openly addressing the real-world complexities and limitations encountered.
The Gossip
Agentic Aspiration & Application
Many commenters expressed curiosity about the concrete output of such parallel agent systems, questioning where the "great software" developed with AI agents is. The author clarified that his current work is internal to Snowflake, but others shared personal projects and internal tools, explaining that public releases are often deterred by strong criticism against AI-generated code, leading much of the current output to remain internal or personal.
Contextual Conundrums & Limits
A significant theme revolved around the challenges of maintaining consistent context across multiple agents, with one commenter noting "keeping their context from drifting" as a bottleneck. The author's stated limit of 8 agents due to cognitive load resonated with discussions in leading AI labs, indicating that scaling beyond a certain human-to-agent ratio requires fundamentally new tooling and improved methods for context synchronization.
Pragmatic Pricing & Preferences
Commenters raised practical concerns, including the likely necessity of top-tier AI model subscriptions for such intensive use, reflecting the financial implications of high-volume token consumption. Despite these considerations, there was clear enthusiasm for the approach, with users expressing appreciation for the system's bootstrap method and advocating for similar tmux and Claude setups in their own development workflows.