Automated AI Agent Experiments

I built an AI team that works completely autonomously—and it's wild to watch. 🤖

Not just automation. True autonomy.

These agents literally sit at computers using git/bash/editors, making decisions like humans do. And here's the thing: this isn't just for developers. Think business analysts, data analysts, content writers, researchers—any role that only needs a computer and decision thinking.

🧠 The Team:

👨‍💻 Senior Dev Agent - Lives on my VM, codes features from scratch, creates PRs with full context, enhance code based on code review

🧪 QA Agent - Tests the live deployed app, writes comprehensive e2e tests, catches bugs before users do

💬 Code Review on GitHub - Auto PR reviews

💬 @claude on GitHub - Tag it anywhere in issues/PRs, it responds instantly and pushes fixes

🔧 The Stack powering them:

📋 Task Master AI - Orchestrates work, manages dependencies, keeps everyone aligned

🎭 Playwright - Browser automation for realistic testing

🛡️ Hooks - Custom safety rules so agents can't break production

▲ Vercel - Auto-deploys preview environments for every PR, mock develop or QA envirenments

🔍 GitHub Actions - Reviews every line of code, runs CI tests automatically

⚡️ The Flow:

It's not a rigid loop—it's like a real team. Based on task dependencies, agents work in parallel or sequentially. Issues can be raised anytime, fixes happen naturally. Just like human SDLC, but faster, with no breaks or coffee!

Task → Dev codes → PR → Review + CI → Deploy preview → QA tests → E2E tests → Regression tests → Merge ✅

Is this automation? Sure, I still set up the VM, prepare tasks, and merge PRs like a project manager. But that's not the point.

✨ The breakthrough is **autonomy**—agents that make decisions, not follow scripts. I can review their PRs, check deployments, of cuz, still can work on code alongside them. It's like managing a team that never sleeps.

✨ What I learned:

The hardest part wasn't the tech setup. It was learning to treat agents as **decision engines**, not deterministic workflows. How to give them goals, not steps. How to let them adapt, not just execute.

That's what makes them effective.

🔗 Watch them work: https://kimwwk.github.io/ohmydoc-using-claude-code-agent/

🔗 Project repo: https://github.com/kimwwk/ohmydoc-using-claude-code-agent

What could your AI team build while you sleep? 🌙

#AI #AgenticAI #Autonomy #FutureOfWork #ClaudeAI