The advice for the last decade was to guard your focus and avoid context switching, because a person can only hold one hard problem at a time. That part still holds. It stops holding the moment an AI assistant is doing the typing, because then you are sitting idle waiting on output that needs none of your attention. That is the argument for learning to run coding agents in parallel.
I have spent 14 years on WordPress and WooCommerce work, long enough to go from dragging files into an FTP client to running proper CI/CD pipelines. Agentic AI has been a bigger jump than either of those. I used to sit and wait for a complex refactor to land before I picked up the next ticket. These days I start three agents and go make coffee, and by the time I sit back down my job is reviewing rather than writing.
The case for parallelism
Typing speed has not been the bottleneck for years. Waiting is. Point Claude Code or Cursor’s agent mode at anything non-trivial and you get a five to twenty minute window where it reads the codebase, thinks, and runs tests. Watching that scroll past is dead time. Run one agent at a time and the whole project moves at the speed of a single LLM response, which is an odd thing to accept when nothing is stopping you from running coding agents in parallel.
I have written before about AI coding agent context. That post was about what the model knows. This one is about what you can hold in your own head while three of them are running, which needs a system of its own. The closest thing I have to a mental model is supervising a few junior devs at the same time.
The four steps I follow
This is what survived a few hundred PRs of trial and error. None of it is about looking busy. It is about keeping enough structure that several running agents do not turn into several half-finished branches.
- Sort by value and effort. Map the tasks somewhere, Todoist or Notion, then send the high value, high effort ones off first. Those are the big rocks, and they keep an agent occupied longest while you deal with something else.
- Spend real time in Plan Mode. Give it 15 minutes before the agent touches a line of PHP. Settle the architecture, name the edge cases, spell out what the agent is not allowed to do. A task planned that way usually runs to the end without you stepping in.
- Drive it from the CLI. A GUI is fine for small edits and gets in the way for the heavy lifting. I use the Claude Code CLI because it lives in the terminal and picks up my shell setup without any fiddling.
- Keep every agent visible. I split the screen in the Warp terminal and give each pane its own agent, either a separate repo or a separate branch of the same project.
Firing a batch from the shell
On a large WooCommerce build you often have several plugins needing the same refactor, deprecated CRUD methods being the usual suspect. Doing those one at a time is a waste when the trigger is scriptable. The bash below is stripped down, but it is the shape I use to start coding agents in parallel across a set of directories.
# Example: Triggering Claude Code across multiple plugin directories
for plugin in ./wp-content/plugins/bbioon-*; do
(
cd \"$plugin\"
echo \"Refactoring $plugin...\"
claude -p \"Refactor all deprecated wc_get_product calls to the new Data Store API. Ensure you run tests after.\"
) &
done
wait
echo \"All agents finished their first pass.\"In a terminal with split panes you can watch each refactor move as it happens. My guide on effective AI programming covers the habits around this.
Where this goes wrong: race conditions and state
I have collected a few scars here. The worst was two agents inside the same database migration file: one renamed the columns while the other added an index to the names that no longer existed. Untangling it took longer than writing the migration by hand would have. So the rule I hold to now is isolation. Two agents never touch the same file or overlapping logic at once. Separate branches, and I merge them by hand after reading the diff.
If getting this set up is eating your week, this is the kind of work I take on. I have been on WordPress since the 4.x days.
What this changes
The point is not volume. Running coding agents in parallel buys you the time to work like a lead engineer, reading and shaping output rather than producing every line yourself. Add one extra agent to your next session and see whether waiting still eats most of the day.
