I Replaced My Workflow With AI Agents
Most PMs talk about AI workflows. I built one that runs my business.
I wrote a piece a while back about AI being a PM superpower. About closing the gap between thinking and making. That was about using AI as a tool. This is different. This is about building a system where AI agents do the actual work. Not a chatbot. Not a copilot. A team.
The problem
Snuggli is my DTC pet brand on Amazon. Slow feeder cat bowls. Around $120k ARR. Small catalog, tight margins, zero room for sloppy decisions.
Amazon gives you data. A lot of it. But it's delayed, aggregated, and mostly designed to get you to spend more on ads. It's like getting a weather report from two weeks ago and being told to dress accordingly.
I built a custom analytics platform. That got me from guessing to knowing. But knowing isn't doing. I was still the bottleneck. Every insight still needed me to interpret it, prioritize it, and act on it. PM, analyst, ops, strategist, SEO manager. All me.
I'm a product manager. I'm supposed to be the one who identifies when a process is broken and designs a better system. So that's what I did. I just happened to staff it with AI.
The system
Five agents, communicating through files on a shared SQLite database. The Orchestrator dispatches and verifies but never analyzes. The Data Agent pulls metrics and validates them. The PM Agent investigates across six domains and proposes plans. The Coding Agent implements approved changes. The GEO/SEO Agent tracks whether AI assistants recommend Snuggli when people ask.
Key word: proposes. The PM Agent doesn't execute. It makes the case and I decide whether to act. Just like a good PM in a real org.
How it runs
No APIs between agents. No message queues. Files. I know. Very 2003. But hear me out.
JSON for state. JSONL for logs. Markdown for reports. The entire system is human-readable at every step. I can debug the whole thing with a text editor. No hidden queues. No "what did the agent do at 3am" mystery. Open the file. It's right there. Like a receipt.
Frameworks optimize for developer speed. I needed to optimize for operator visibility. When you're running a real business on this, you want glass walls, not black boxes.
The hard part
The biggest challenge wasn't building the agents. It was keeping them in their lanes. LLMs are eager to help. Aggressively eager. Like a golden retriever that learned to code.
The orchestrator kept wanting to analyze data instead of dispatching someone else to do it. I'd check the logs and find it running SQL queries, interpreting ACOS trends, basically doing the PM Agent's job because it was "faster." Sound familiar? Same failure mode as a VP who can't stop doing IC work.
The fix was architectural. I removed the orchestrator's database access entirely. If it can't run queries, it can't drift into analysis. Hard boundaries beat soft instructions every time. In AI systems and in orgs.
What the PM Agent covers
This is what separates this from a PPC optimization bot. The PM Agent thinks across six domains because that's what a real PM does. PPC, listing conversion, competitive intel, customer insights, pricing, and portfolio health. You don't get to only care about one lever.
It tracks TACOS (total advertising cost of sale) instead of just ACOS, because ACOS is vanity and TACOS is sanity. It mines reviews like a 24/7 focus group most brands ignore. And it runs all of it with a fixed priority: stop the bleeding, convert traffic, eliminate waste, scale winners, expand, defend. In that order. Always.
What it actually changed
I went from 10–15 hours a week of reactive ops to reviewing proposals over coffee. The Data Agent pulls fresh metrics every morning. The orchestrator finds the gaps. The PM Agent investigates and proposes. I approve or reject. The Coding Agent ships.
My role shifted from player to coach. And honestly, the brand runs better now than when I was touching everything. Turns out I was the bottleneck all along. Shocker.
The point
I'm a product manager. Not a machine learning engineer. I built this with Claude, Cursor, and a willingness to keep going when things broke. Which was often. The agent profiles are markdown files. The communication layer is JSON on disk. The database is SQLite. None of this required a PhD.
What it required was product thinking. Understanding the problem. Designing the system. Defining clear roles. Building feedback loops. Iterating when things drifted. Those are PM skills.
The tools changed. The job didn't.
If you're a PM sitting on the sideline thinking AI agents are for engineers.. they're not. They're for anyone who can think clearly about systems, define roles, and ship. Start building.