The Claude Code Subagents I'm Actually Using Right Now

For a while, I was doing what most people do with Claude Code - one window, one prompt, and just asking it to review, debug, and fix everything myself. It works fine until it doesn't. The moment your codebase gets even a little complex, one generalist AI trying to hold context for code review, debugging, security, and architecture all at once starts missing things. Not because it's a bad model - because it's doing ten jobs at the same time.

That's what pushed me toward the awesome-claude-code-subagents repo - a free, open collection of 100+ specialist agents you can drop straight into Claude Code. Instead of one AI juggling everything, you get narrow, single-purpose agents that only do one job - and are noticeably better at it. Here are the ones I'm actually running in my own projects right now.

debugger is the one that sold me on this whole approach. I pointed it at a rate-limiting issue in one of my services, and instead of a vague "this could be a race condition somewhere" answer, it pulled the actual log files, counted the exact failures (106 in one night, across 22 of 27 API keys), traced it to a specific commit, and confirmed the root cause with a git diff. That's a completely different tier of output than a plain prompt gives you.

code-reviewer is my day-to-day one - I run it after almost every meaningful change. It's read-only by design, so it won't touch your files, it just flags issues by severity with file and line references, the way a senior engineer would in a PR review.

security-auditor catches the stuff a general review pass tends to skip - auth flaws, exposed secrets, insecure dependency patterns. I run this before anything touches production, especially services handling API keys or user data.

architect-reviewer is useful less often but hits harder when it matters - it looks at whether a design decision actually holds up structurally, not just whether the code runs.

Beyond those four, a few others worth knowing about depending on your stack: devops-engineer and kubernetes-specialist if you're deploying anything containerized, database-optimizer if your queries are the bottleneck, api-designer if you're building anything REST or GraphQL-facing, performance-engineer for anything that needs to be fast under load, test-automator if your test coverage has gaps, refactoring-specialist for cleaning up legacy code without breaking it, and documentation-engineer if - like most of us - your docs are perpetually behind your code.

None of this replaces actually understanding your own codebase. But the difference between asking one AI to be your whole team versus handing each task to something built specifically for that task is real, and it shows up in the specificity of what comes back. If you're already using Claude Code, this repo is free, vendor-neutral, and takes about two minutes to install.