Skip to content

pxl-stack

The agent harness PXL builds with: skill files, subagents and parallel worktrees on Claude Code.

Internal tool

The problem

Why this exists.

AI coding agents are fast and unreliable in the same breath. Left alone they drift from conventions, repeat mistakes and produce changes nobody has read.

The speed is only worth having if the output holds up in production.

What it does

How it works.

Skill files

Written conventions for each kind of task, so agents work the way the studio works.

Subagents

Narrow agents for research, implementation and review, each with a focused brief.

Parallel worktrees

Independent changes run side by side in separate git worktrees, so they don't collide.

Human review

Nothing merges without a person reading the diff. Agents draft; the studio ships.

How it's built

Principles before features.

On Claude Code

Built around Claude Code's skills, subagents and worktree support.

Used on every build

Client projects and Labs products both run through it, which is how a Pilot Sprint lands in weeks.

Status

Internal tool

What's next

  • Deciding whether to open-source it.
  • Writing up the workflow in the Studio Log.

Want this for your business?

Everything we learn building Labs products goes into client work.