Working on many things in parallel on the surface is very overwhelming.
So we need to start by creating a slow and intentional process for shipping high quality features (i.e. brainstorming documents, planning documents, todos, triage, multi-agent reviews, etc). Create your own, or use plugins like compound engineering/gsd/superpowers.
Compound engineering for example can take many minutes between each prompt as it explores and thinks. It creates great output (given strong input) at the cost of time, like any person would.
Once you have a process you like, it should be the equivalent of you pair coding with a better version of yourself.
Pair coding with one person at a time is not scalable.. I.e. trying to watch the changes and pair code with two people writing different features at the same time would be a nightmare.. and the same can be true with pair coding with a few agents in parallel.
So to leverage worktrees you need to shift your perspective of shipping a single feature, to managing the outcomes of many engineers.
Imagine each worktree is an engineer on your team, assign work the same way (i.e. no two worktrees should be working on exactly the same feature), then simply answer their questions/help them test their changes/provide feedback.
You only review code when the worktree agent has reviewed their own code enough times that they (Claude) are happy with the result and submit a PR. Then you review the code, just like any other person on your team. Ask for changes and back to testing.
AI makes code is cheap, your time is still valuable, so figuring out how to scale yourself is always going to be better than a tool that tries to scale for you.