March 2, 2026
Eight interns, one tmux—what could go wrong?
Parallel coding agents with tmux and Markdown specs
One dev, eight AI interns, and a notes file—fans hype, skeptics ask: where’s the app
TLDR: One developer runs multiple coding bots using simple notes and commands, claiming 4–8 is the workable limit. The crowd splits between admiration for the no-frills system and demands for real results, with warnings about “context drift” and costs shaping a bigger debate about the ceiling of AI-assisted coding.
A solo dev claims they’re wrangling a squad of coding “agents” (think helpful bots) with plain tools: tmux (a split-screen terminal), Markdown design docs, and simple slash commands. Each window plays a role—Planner, Worker, PM—and every feature gets a tidy spec with problem, options, final plan, and checks. Commits even tag the spec, so the changelog writes itself. The catch: they say 4–8 agents is the sweet spot; beyond that, quality dips.
The comments? Spicy. The top vibe is “show me the goods”—one critic demands proof of real apps, not just “AI influencer shovels.” On the hype side, fans love the “bootstrap from one Markdown file” move: low-friction, no fancy orchestration, just notes that act like a pocket project manager. Practitioners get grim: context drift turns multiple agents into competing realities—“by hour two, each pane believes a different truth.” A researcher-y voice says the 8-agent ceiling tracks with labs and points to this deep dive, hinting we need new tools to go past a 1 human : 10 bots ratio. Meanwhile, a wallet-check lands hard: does this need a top-tier AI subscription?
Memes about “tmux sweatshops” and “eight unpaid interns” flew, but the core fight is clear: elegant process vs. productivity cosplay.
Key Points
- •The workflow coordinates parallel coding agents using tmux, Markdown, bash aliases, and six slash commands without complex orchestrators.
- •Work is organized around Feature Design (FD) Markdown specs that include problem, alternatives, chosen solution, implementation plan, and verification steps.
- •FDs are numbered, tracked in docs/features/, and move through eight lifecycle stages managed by slash commands.
- •Every commit is tied to an FD, enabling automatic changelog accumulation as features are completed.
- •An example FD details a multi-label document classification feature using LLMs, few-shot examples, and specific code/SQL changes, with statuses tracked in FEATURE_INDEX.md.