Parallel coding agents with tmux and Markdown specs

Original Article ↗ Hacker News Discussion ↗

Perceived productivity vs. “where’s the software?”

  • Some are skeptical: if multi-agent setups are so productive, they ask why we don’t see lots of clearly “great” AI-built software.
  • Others argue it’s early: tools only got good recently, feedback cycles for software quality are long, and much of the output is internal tools or personal projects.
  • Several note that AI mostly increases volume of “mundane” but useful software, not necessarily greatness; average quality may even drop as volume rises.

Reported use cases and concrete wins

  • Many describe internal or personal tools: finance and chat apps, window compositors, status dashboards, automation scripts, filesystem/Ansible helpers, browser automation, CI tooling, ZFS backends, games, and reverse‑engineered docs for legacy systems.
  • One user claims a ~20%+ reduction in PR time-to-merge via parallel review agents, but this is challenged as over-extrapolated and not business‑proven.
  • Others mention multi-agent code review at large companies and production-grade products built on agentic coding.

Orchestration patterns: tmux, specs, worktrees

  • Common pattern: one markdown spec per agent/pane; agents work in parallel against git worktrees or repo copies to avoid clashes.
  • Some prefer 2–3 focused agents (backend/frontend/tests) over 6–8, citing merge conflicts and cognitive overhead.
  • Others build higher-level “factory” abstractions with a supervisor agent that decomposes work, spawns workers, and manages worktrees/merges.

Context management and documentation

  • Major challenge is context drift across sessions; solutions include:
    • Per-agent spec docs plus an orchestration doc.
    • Tools like agent-doc, Beads, or entity-centered “NERDs” documents.
    • PROJECT.md / SPEC.md to record direction, key decisions, and avoid scope creep.

Costs, quotas, and optimization

  • Parallel agents rapidly exhaust top-tier subscriptions; several hit weekly limits in 3–4 days.
  • Strategies: mix cheaper models, “oracle” agents for code questions, cheap supervisors to detect spec gaps, and strong planning/checkpointing to reduce wasted “thinking” tokens.

Quality, validation, and safety

  • Strong emphasis on tests, verification commands, and sometimes separate reviewer agents.
  • Concern about agents bypassing deny lists; some invert the model and require “proof of safety” (explicit intents, path checks, diffs, tests) before any action.
  • Long-term impact on maintainability and design quality is seen as unclear and unmeasured.