The WorkChapter 04 / 05

Chapter 04 / 2026 / Agents

Software Factory

A team of AI coding agents that takes a software job, plans it into pieces and builds the pieces at the same time. One agent leads and the others build, each in its own copy of the code. Scripts test every piece before it is merged.

What it is

One agent is a bottleneck

One agent, one file at a time. That is slow.

A team of AI coding agents that takes a software job, plans it into pieces and builds the pieces at the same time.
The team at work in Herdr: the Lead and four builders listed down the side, and one builder's tests passing.
Who it is forAnyone who uses AI agents to write software and wants the work done sooner without paying more or getting worse code.

How it works

A Lead, lanes and checks

Give each builder its own files, and let scripts do the checking.

1. Plan with the personThe Lead holds the plan, a numbered list of issues agreed with the person, before any builder starts.
2. Split into lanesThe builders' room: each builder at a desk of its own along its own lane, so that no two of them touch the same files.
3. Build side by sideThe team at work in Herdr: the builders listed down the side, and one builder's code and tests in its own worktree.
4. Land and reviewEach lane joins the main line by its exact commit, past the reviewer, and then the tests run on it.
The team at work in Herdr: the Lead and four builders listed down the side, and one builder's tests passing.

What went wrong

The run that stalled

The code was fine. The rules around it stalled the run.

On the first real-world benchmark, one agent alone finished in 19 minutes.
The team was stopped at 48, with 10 of 14 tickets done.

What it measured

Faster at the same quality

Faster every time, and no worse.

−36% to −61%The team needed between 39% and 64% of the time one agent took alone.
3 of 3Each time, the team's code passed at least as many unseen tests as the single agent's.
−24% to +8%From a quarter cheaper to 8% dearer than one agent, inside the 15% limit set at the start.

My part and the stack

My partDesigned the team and its rules, wrote the scripts and the planning playbook, ran the benchmarks against one agent and wrote the public build log.
The results table from the project page: time, cost and hidden-test scores for the team against one agent alone.