Running AI coding agents in parallel helped me build several features at once for a sports betting news publisher: a tipping record, a sports calendar and map, racing results, a glossary, a settings panel, a bet slip, dark mode, caching and an admin command centre. Much of it was written by AI coding agents, several running at the same time, with me directing, reviewing and deciding.
Running one agent well is a skill. Running many is mostly a coordination problem, and the problems are the same ones a human team has, only faster.
AI coding agents: one writer per area
The first rule, and the one I broke first: every area of the code has exactly one agent allowed to change it at a time. A plugin, a template part, a stylesheet. Two agents fixing "the header" at once will each make a reasonable change, and the second will quietly undo the first.
The awkward areas are the shared ones. The header, the theme's dark mode styles and the global stylesheet are touched by almost every feature. Those got a single owner. Other agents describe the change they need and the owner makes it, the same way a team routes changes to a shared component through the person who maintains it.
One locked way to deploy
The dev server was the shared place where everyone's work met. Early on, agents deployed whenever they finished, and a deploy overwrote another agent's deploy that had landed minutes earlier. Each agent reported success, correctly, about its own work. Together they had lost a change.
The fix was one deploy script that every agent must use. It takes a lock so only one deploy runs at a time, and it deploys what is committed to the main branch rather than whatever happens to be in someone's working copy. An agent that finds the lock held waits. This is not clever. It is the same discipline as giving deployment config an owner, applied to a team that types very quickly.
Measure the reference, do not eyeball it
A lot of the work was "make this look like that": match the settings panel's header, match the original logo, line the rail up with the header. Asked to match something by eye, an agent produces something close, and close accumulates. Three components each "about right" give you a page that is visibly off.
What worked was measuring first. Read the computed sizes, spacing and colours of the reference, write them down, then build to the numbers. The animated text logo was matched to the original image logo by laying one over the other and comparing pixels. Better still, where two things should be identical, use the same component. The bet slip's close button is now the settings panel's close button, so they cannot drift.
A claim is not a result
Agents report success confidently. Sometimes the change is perfect. Sometimes it works at one width, in light mode, with default settings, and nowhere else. The report reads the same either way.
So nothing counted as done until there was a screenshot of it working, and not just one screenshot. My standard set for an interface change:
- phone and desktop widths;
- light and dark mode;
- reduced motion on;
- the dyslexia-friendly font on, which changes text metrics everywhere;
- a cached page as an anonymous reader, as well as a signed-in view.
The dyslexia font check alone found a bet slip tab that clipped its second line and a price line that cut off the return. Both had been reported as finished.
Write the commit message for a person
With many agents committing, the history is the only reliable record of what changed. I asked for commit messages that say what a reader of the site would notice, in plain English: "matches with no known start time say TBC" rather than "fix null time handling". When something broke, that history made it quick to find which change to look at, and it doubled as a changelog for the client.
What I would set up on day one next time
- A written map of who owns which area, including the shared files.
- A locked deploy script before the first deploy, not after the second collision.
- A rule that "match this" means "measure this first".
- A standard screenshot set that every interface change must pass.
- Plain-English commit messages.
The agents made the build far faster than I could have managed alone. The speed was real. So was the need for the same boring coordination any team needs, set up before it is missed rather than after.