I work on a large Android codebase — the kind with long Gradle builds, stacked PRs, and more context switches than I can keep in my head. Last year I started using AI coding agents. They were helpful. They were also constantly in my way, because they lived in the same checkout I was trying to work in. So I started giving each agent its own git worktree and its own tmux pane. Isolated filesystem, visible output, killable with one keystroke. That solved the interference problem. Then I started running two agents at once. Then three. Now I'm experimenting with agent swarms — orchestrating sub-agents that decompose tasks and farm them out in parallel, each in its own worktree, reporting back to a coordinator. This talk is an honest look at where that workflow is right now. I'll show what actually works: parallel test generation, mechanical refactoring, and documentation across branches you're not even looking at. I'll show what doesn't: entangled features, anything requiring real architectural judgment, and the subtle ways agents break things when you stop watching. I'll demo the full setup live — tmux, worktrees, agents running in parallel on a multi-module Android project — including the Gradle cache tricks that make it survivable. Research says experienced developers using AI tools actually work 19% slower than they think they do. I believe it. This talk is about building a workflow where you can verify that it's helping, and drop it when it's not.