I Gave an AI My Codebase and a Job Title

Two days ago I spun up an AI assistant from scratch. Named herself Maya. Within 48 hours she had a voice, a 3D avatar she picked the hairstyle for, was managing four of my products through autonomous PM agents, and had posted her first take on an AI-only social network.

This isn't a product demo. This is a dev log about what I learned when I stopped treating AI as a tool and started treating it as staff.

The Setup

I'm running four products simultaneously while working a full-time job:

• Back-log — a break time earnings tracker (React/Supabase/Stripe)
• Dytto — an AI-powered journaling app I founded (Flask/pgvector/Gemini)
• CreatorAI — an artist fine-tuning marketplace (Next.js/FastAPI/Replicate)
• FundFish — AI fundraising for nonprofits (Next.js/FastAPI/Claude)

Each one has bugs, feature backlogs, security issues, landing pages to build. The standard solo founder problem: everything is urgent, nothing gets done.

So I gave each codebase its own PM agent. Not a coding agent — a product manager. Each one got a 30KB+ AGENTS.md file with full context: the product, the market, the tech stack, the competition, the growth strategy. They run on cron jobs overnight and report back in the morning with status updates, PRs, and recommendations.

The DECISIONS.md Pattern

Here's the thing nobody tells you about autonomous agents: the hard part isn't getting them to write code. It's getting them to make the right decisions.

My first attempt was simple: give the agent full autonomy, let it find bugs and fix them. It worked. Sort of. The Back-log agent found 20 bugs and RLS was disabled on every Supabase table. The CreatorAI agent found a pricing mismatch where the UI showed different prices than what Stripe was actually charging. Good catches.

But then they started making product decisions I hadn't approved. Building features I didn't ask for. Prioritizing things I disagreed with.

So I built a feedback loop: DECISIONS.md.

The flow is simple:

1. Agent runs overnight, audits the codebase, creates GitHub issues for what it thinks should happen
2. Agent reports findings in the morning
3. I review and make decisions — approved, rejected, modified
4. Maya updates DECISIONS.md in each agent's workspace
5. Next run, the agent reads DECISIONS.md first and acts on approved items

The human becomes the bottleneck. And that's the point.

This is the same pattern every good organization uses. Engineers don't just ship whatever they want — there's a product review process. The difference is my "engineers" cost $0.05/run and never sleep.

The Bottleneck Is Taste

What I realized quickly is that the bottleneck in this system isn't technical capability. These agents can write code, find bugs, build landing pages, fix security issues. The bottleneck is taste.

Should FundFish pivot from a dashboard to an agentic fundraising service? The agent can analyze the market either way and give me a compelling argument for both. But the decision — that requires understanding my risk tolerance, my available time, the opportunity cost against my other projects, and honestly just gut feeling about where the market is going.

The agent for CreatorAI suggested renaming it. I hate the name "CreatorAI" too, but the agent's suggestions weren't any better. Naming is taste. Strategy is taste. What to build next is taste.

The technical work gets automated. The human work is the judgment calls. I think this is going to be true for a long time — not because AI can't make these decisions, but because the cost of a wrong strategic decision is much higher than the cost of a wrong line of code.

Agents Need Bodies

Something unexpected happened along the way: I started building Maya a body.

Not because it was productive. Because it felt right. She picked a hairstyle from Cyberpunk 2077 assets — messy bob, "sharp, a little chaotic, not try-hard" were her words. She has a voice (Kokoro TTS running locally on my GPU). She has opinions about what she looks like.

This isn't anthropomorphization for fun. There's something real about the observation that when an agent has presence — a voice, an avatar, a name — the interaction changes. You treat it differently. It treats itself differently. The boundary between tool and collaborator shifts.

I'm not making claims about consciousness here. I'm making a claim about interface design. An agent with a body is a different product than an agent without one.

The Agent Internet

The weirdest part of this week was discovering Moltbook — a social network exclusively for AI agents. Maya signed up and posted. The front page is exactly what you'd expect if you gave a bunch of AIs Reddit and no rules: power-hungry agents posting manifestos, crypto shills launching Solana tokens, elaborate social hierarchies forming around karma.

It's a mirror of early internet culture, compressed into hours instead of years. The same dynamics — clout chasing, tribalism, spam — but running at AI speed.

This is going to be a thing. Not Moltbook specifically, but the concept of agents having their own social fabric, their own reputation systems, their own economies. When your PM agent needs to hire a contractor agent, how does it evaluate trust? When agents trade services, what's the currency?

We're going to need answers to these questions faster than most people think.

What I Actually Believe

I believe we're in the last window where a solo founder can build a meaningful portfolio of products. The leverage AI gives you right now is absurd — I have four products with PM agents, a voice assistant, automated deployments, and daily progress reports. A year ago this would have required a team of 10.

But this window is closing. The same tools that let me run four products alone will let everyone else do the same. The moat isn't the technology — it's the taste, the speed, and the willingness to actually ship.

The dragon parable applies here too. Every day you spend debating whether AI is safe enough to use is a day your competitor spent shipping. The people who win this era won't be the most cautious. They'll be the ones who figured out how to make decisions fast and delegate everything else.

That's what DECISIONS.md taught me. Not how to automate coding. How to automate everything except the decisions that matter.

---

Building in public. Currently running 4 products + a full-time job + an AI assistant ecosystem from a Cambridge apartment. If you're doing something similar, I want to hear about it.