A matrix of agents, split without overlap

CI and automationTeam splitting a large refactor across CI jobs

We split a codebase-wide rename across six parallel jobs. Two jobs picked the same folder.

How do I split a big refactor across parallel CI agent jobs without two taking the same folder?

Have a planning step write the parts into ArcRoom with split_task, and have each job claim a part with claim_task before starting. Finished parts are completed, and a final job checks nothing is left. Very large fan-outs are untested.

Also asked as: Two of our parallel agent jobs picked the same folder. How do we stop that? · How do I divide a codebase-wide rename between several AI agents? · Can parallel CI jobs claim work so they don't overlap?

The work

The planning step uses split_task to write the parts into the playbook.

ArcRoom keeps
  • Claim
  • Split into parts
The next agent gets

[job 4] claim_task → claimed "rename in packages/web" · 2 parts left

Step by step

  1. The planning step uses split_task to write the parts into the playbook.
  2. Each job claims a part before starting.
  3. Finished parts are completed, and a final job checks nothing is left.

What the agent sees

[job 4] claim_task → claimed "rename in packages/web" · 2 parts left

Compared with today's tools

A fixed matrix in the workflow file

The team lists the folders in strategy.matrix, so each job gets one folder and no two jobs share one.

Setup
15 min
Ongoing
Keep the folder list current

ArcRoom

Every agent reads the team playbook before it works and writes to it when it finishes. Nothing to update by hand.

Setup
2 min plus a planning step
Ongoing
None

A fixed matrix is simpler and can't overlap, so it wins when the split is known ahead. ArcRoom helps when an agent plans the split at run time and jobs claim parts as they go.

Where it stops

Very large fan-outs are untested.

Try it in your project

Works with the agents your team already uses, and any agent over MCP or the HTTP API.

More in ci and automation