Last week, my screen flickered alive in the dim glow of my home office — Copilot CLI churning through a dusty codebase I’d abandoned for months.
AI subagents. That’s the phrase buzzing in my head now, after feeding this beast a vague prompt: “analyze the following codebase and report improvements and possible bugs.” Vague on purpose — leaves room for gems amid the garbage. It spat out a dozen items. Solid start.
I pushed further. “Create a GitHub issue for each, with labels, including priority.” Boom — issues born. But three flagged phantom libraries, versions newer than its training data. Wrong. Closed as won’t fix. Classic AI hallucination, but hey, progress.
Triage time. Half felt fishy, half gold. Four survivors. Now, the fun part: unleash subagents.
Why Bother with AI Subagents?
Here’s the thing — newbies botch this. They overfeed agents every scrap of context, hoping for autonomy. Waste. Triage already arms you with details. Subagents thrive on isolation, firing up fresh contexts like independent rockets.
My prompt? Crystal. For each issue — X, Y, Z — launch a sub-agent to: fetch via gh tool, read description, spin a git worktree branch, implement fix or feature, add tests if needed (all must pass), semantic commit, push, PR with pattern “fix-issue-[title]-by-ai”.
For each issue X, Y, & Z, I want you to launch a sub-agent that: - Fetch the issue using the gh tool - Read its description - Create a dedicated branch using the git worktree command - Implement the feature or fix the issue - If the feature/issue warrants it, create a test or tests around it - All tests must pass before you continue - Commit using a semantic commit - Push it on its own branch to GitHub - Create a PR with this branch, using the following naming pattern
Two things of note.
Auth via gh in terminal — Copilot grabs full perms. Git worktrees? Genius hack. One repo, multiple folders, no toe-stepping. Like cloning elves into separate workshops, each hammering their anvil without chaos.
Parallelism hits first. Sequential research, sure — but fixes fly simultaneous. Main win? Context purity. No token bloat from prior chats, RAG noise, or system prompts bleeding over. Each agent: pristine slate. Engineer’s dream — efficiency without excess.
And look — Claude Code’s optimizing contexts now. Smart. But subagents? They force the discipline.
Is Git Worktree the Unsung Hero Here?
Knew of worktrees forever. Never needed ‘em. Till now.
A git repository can support multiple working trees, allowing you to check out more than one branch at a time. Picture your main codebase as a bustling city square — worktrees sprout neighborhoods, each alive with its branch’s changes. No merge hell. git worktree add ../fix-issue-X fix-issue-X — done.
Subagents dance in their silos. Copilot CLI, hooked to GitHub MCP (read-only default — flip with gh), executes flawlessly.
But — reality check. Agents won’t query back much. Design upstream sharp, or they veer. Review everything, like a foreman eyeing junior devs’ output.
Difference? Minutes, not days. No coffee runs. No egos. We’re orchestrating digital minions, not mentoring humans.
My unique twist: this echoes the Ford assembly line’s dawn. Back then, mass production shattered craft guilds — one man per task, parallel belts whirring. Today, AI subagents birth code factories. Bold prediction? By 2026, solos and startups run full AI dev squads, churning MVPs overnight. Hype? Nah — I saw PRs land clean in under an hour.
Corporate spin screams “autonomy,” but it’s delegation 2.0. You’re still boss — just with godlike speed.
Can AI Subagents Replace Your Junior Devs?
Short answer: not yet. They’re tireless, but blind to nuance without your blueprint.
Pros explode. Parallel fixes — four issues, four branches, tests green, PRs up. Context isolation slashes errors. Like herding cats into laser-focused lions.
Cons? Hallucinations linger (those ghost versions). No creative leaps sans prompt. And review load spikes — but faster than manual coding.
We’re shifting. Companies ditch junior training for agent wrangling? Makes sense at scale — why onboard when AIs scale infinitely? But lose that human spark? Tradeoff.
Tested on a mid-sized Node repo. Two bugs squashed, one refactor polished, one feature stubbed — all passing CI. Wonderment.
Deeper: imagine scaling. Feed 50 issues? Subagent swarm. Open source maintainers, rejoice — backlog apocalypse averted.
Setup hurdles? Terminal auth. Worktree savvy. Prompt tuning. But once humming — electric.
How Do You Launch AI Subagents in Your Own Repo?
Step one: gh auth login. Terminal only.
Prompt as above, tweak for your stack. Java? Add system: “Expert Java architect, 20+ years.”
Watch terminals multiply — or chain in one, subagents sequential if needed.
Fun fact — polluted main context? Subagents dodge it entirely.
This isn’t toy. It’s platform shift. AI as dev OS layer, subagents as apps.
Energized yet?
🧬 Related Insights
- Read more: Why Your ADHD Med Tracker App Vanishes by Thursday
- Read more: Anthropic’s Epic Oops: Claude Code’s Source Spilled Wide Open
Frequently Asked Questions
What are AI subagents in coding?
Tiny autonomous AIs spun from a main agent, each tackling one task with isolated context — like mini-devs fixing code in parallel.
How do git worktrees enable AI subagents?
They create separate folders for branches in one repo, letting multiple agents work without conflicts — perfect for parallel fixes.
Will AI subagents fix my entire backlog?
They crush triaged issues fast, but human oversight prevents hallucinations; scale with sharp prompts for best results.
