parsak

My co-founder Tyler Brown and I have been building our product for 6 months. The co-working space that Tyler founded that we work out of houses 300 founders that we've gleaned agentic coding tips and tricks from.

Neither of us came from traditional SWE backgrounds. Tyler was a film production major. I did informatics. Our codebase is a 300k line Next.js monorepo and at any given time we have 3-6 AI coding agents running in parallel across git worktrees.

Every feature follows the same four-phase pipeline, enforced with custom Claude Code slash commands:

1. /discussion - have an actual back-and-forth with the agent about the codebase. Spawns specialized subagents (codebase-explorer, pattern-finder) to map the territory. No suggestions, no critiques, just: what exists, where it lives, how it works. This is the rabbit hole loop. Each answer generates new questions until you actually understand what you're building on top of.

2. /plan - creates a structured plan with codebase analysis, external research, pseudocode, file references, task list. Then a plan-reviewer subagent auto-reviews it in a loop until suggestions become redundant. Rules: no backwards compatibility layers, no aspirations (only instructions), no open questions. We score every plan 1-10 for one-pass implementation confidence.

3. /implement - breaks the plan into parallelizable chunks, spawns implementer subagents. After initial implementation, Codex runs as a subagent inside Claude Code in a loop with 'codex review --branch main' until there are no bugs. Two models reviewing each other catches what self-review misses.

4. Human review. Single responsibility, proper scoping, no anti-patterns. Refactor commands score code against our actual codebase patterns (target: 9.8/10). If something's wrong, go back to /discussion, not /implement. Helps us find "hot spots", code smells, and general refactor opportunities.

The biggest lesson: the fix for bad AI-generated code is almost never "try implementing again." It's "we didn't understand something well enough." Go back to the discussion phase.

Wrote up the full workflow with details on the death loop, PR criteria, and tooling on my personal blog, can share if folks are interested.

All Claude Code commands and agents are open source: https://github.com/Dcouple-Inc/Pane/tree/main/.claude/commands

I built Pane using this workflow over the last month or so and started using it every day a little over a week ago. Each workspace gets its own worktree and session. Takes me less than 30s to go from idea -> discussion now. The repo link above takes you to it.

On a good day I merge 6-8 PRs. Happy to answer questions about the workflow, costs, or tooling for this volume of development.

I work on a 300k line monorepo, just me and my co-founder. At any given moment I have 3-6 CLI agents (Claude Code, Codex, Aider) running simultaneously across git worktrees. The throughput is great. Managing it is not.

Every tool I found was either another agent, an IDE plugin, an abstraction over my CLIs, or doesn't understand worktrees. Conductor is Mac-only and getting buggier. Warp and Ghostty were interesting but not opinionated enough for my worktree-to-PR workflow.

So I built Pane (https://runpane.com). Keyboard-driven desktop app that gives you one interface to monitor and control CLI agents across worktrees. Ships with a command palette, simple shortcuts (ctrl + up or down arrow to switch between each worktree, VS code shortcuts for other basic things). Each worktree gets a run button that auto-generates a script (via Claude Code on first run) to spin up on isolated ports, so I can have every branch hot reloading in its own tab.

Been using it daily since last week. Hard to go back. I fully open sourced it here, so you can ship your own features to Pane, using Pane: https://github.com/Dcouple-Inc/Pane

How are others handling multi-agent workflows?