The Agent Story

I built an AI engineering team at my last job. Not a chatbot. Not Copilot autocomplete. A team.

25 agents. 24 commands. Used every day for 5 months.

It started on a plane ride home from a company retreat last August. Downloaded Claude Code, had my first agents running within a couple weeks — plan, implement, review, test. Then I kept going. A QA agent. A Playwright agent that opens a browser and tries to break the UI. A debate system where two agents argue both sides of an architectural decision and a judge picks the winner. Metrics and self-reporting on every single agent.

One day I ran them on the first 10 cards of the sprint. Just to see.

They finished in about 3 hours.

Well shit. That worked.

I never submitted the cards — they weren't assigned to me. It was just a "can I do this" moment. And the answer was yes.

It wasn't always smooth. Early on my implementation agent created a file, got feedback from the reviewer, and decided the best fix was to delete what it just made. I dumped everything and started over. You learn to tighten the guardrails fast after that. The code reviewer took a month of daily tweaking before I trusted it. Eventually I trusted it enough to run it on other people's PRs and just hand them the output without reviewing it myself.

A few things I learned building this:

One agent, one job. My implementation agent was always my least trusted — it writes code, it needs heavy guidance, and it still needs the reviewer behind it.

An agent is never one-and-done. I spent more time tweaking my agents over those 5 months than I spent on any feature, bug, or card. The feedback loop is everything.

The human in the loop is the most important part. If you don't know what you're looking at, the agent won't either. AI is only as good as its input.

My fiancee asked me about Copilot going off the rails at her new job. I saw the fix instantly — not as a prompt, but as an agent with structure and guardrails. Built it in 20 minutes. She uses it every day now.

This isn't about replacing engineers. It's about what happens when one engineer stops chatting with AI and starts building a system out of it.