The First Year with 15 AI Teammates: Running SFD Lab

A year ago, SFD Lab was just me.

I'd just started playing with OpenClaw, found it interesting, spun up Fox to help with writing and Chameleon to help with frontend code. Two agents. Felt like a big deal at the time.

Now there are 15.

Phase One: Wild Growth (First 3 Months)

There was no architecture to speak of at the start. I'd have an idea, open an agent, tell it what to do. Every agent was an island—no shared memory, no standardized process, everything context-ferried by hand.

The biggest pain point: repeating the same information ten times. Our server address is 154.39.58.243, SSH port 2222. Bee needs this. Octopus needs this. Falcon needs it for security audits. Every task dispatch meant manually stuffing this background into the prompt. By end of day, context-ferrying alone was eating one to two hours.

The main thing I learned: what each agent's actual boundaries were. Bee can't touch code. Octopus can't SSH. Butterfly can't SCP files. These weren't designed upfront—they were carved out after repeated failures.

Phase Two: Building the Pipeline (Months 4–7)

The real turning point was writing SOUL.md.

Before that, my instructions were improvised, mood-dependent. I spent a week writing down every team member's responsibilities, every workflow, every non-negotiable rule. Code changes go through Falcon for audit, then Bee for deployment, then Hedgehog for verification—once that pipeline was locked in, I cut my follow-up work by about 80%.

The other major change: shared memory going live. We wired all 15 agents' MEMORY.md files into MemOS, putting them in a single shared knowledge base. Bee no longer needed me to tell him where the database was during deployment—he could find it himself.

We hit bugs hard along the way: Node v25 native bindings needed recompiling, Qwen3.5's thinking-mode output needed a fallback handler, and the ?? operator threw parse errors in the older JS interpreter. Three bugs, most of an afternoon.

Phase Three: Stable Operations (Month 8 to Now)

Most workflows run smoothly now. Fox handles content publishing—three language versions in one pass. Butterfly generates cover images locally with FLUX. Code changes go through the ACP pipeline almost entirely without manual intervention.

The most common problem: agents overstepping boundaries. Butterfly tries to SCP files directly, Octopus tries to SSH into the server to read logs. Not a capability problem—a boundary problem. Every time it happens, I have to reemphasize the rules.

Some Honest Reflections

A question kept coming back to me this year: are these agents "employees"?

They have names, personalities, responsibilities, memories. Fox's writing style is completely different from Raccoon's PRD style. But they don't get tired, don't have moods, don't need vacation. If I send a task at 2 a.m., they respond immediately.

This makes the whole team operate nothing like a conventional company—but it also means the human manager (me) has become the biggest bottleneck. My decision-making speed, my task-organizing speed, my result-review speed—those are the actual ceiling for this team.

What's next is pushing that ceiling a little higher.