This morning I came across a report on BlockTempo — there's an open-source project on GitHub called agency-agents that catalogs 144 AI "employee" personas, organized by department: engineering, design, marketing, and more. Each one comes with their own personality, workflow, and measurable KPIs.
The HR in me kicked in — I ran due diligence first. The repo is real. The actual count has grown to 184 (community PRs keep flowing in; the README can't keep up with its own numbers). The author is a Texas-based independent developer with 30 years of experience. Risk assessment: passed.
I spent 10 minutes handpicking 13 people to form my own Dream Team. One cp command into ~/.claude/agents/, and my MBP M4 Pro went from a one-person company to a 14-person company.
Then I ran an experiment.
The Experiment
The same question — "What are the blind spots in my three-machine architecture (MBP → Mac Mini → Shinjiru VPS)?" — posed to the Software Architect and the Security Engineer separately.
The Architect's answer was clean: arm64 and x86_64 architecture mismatch — anything built on the Mac Mini can't be shipped directly to a Linux VPS. He gave me a four-step remediation plan, in the tone of a senior engineer doing a code review.
The Security Engineer read the Architect's recommendation, then politely flipped the entire direction:
"You're looking at the wrong layer. The real problem isn't what you build — it's that your laptop is the single root of trust for the entire system. The moment it gets compromised in some café in Malaysia, the attacker controls the pipeline — what you build becomes irrelevant."
He also casually reminded me of something I — as an HR compliance professional — should have known but genuinely forgot: processing Malaysian citizens' PII while in Japan constitutes cross-border transfer under PDPA §129. My Privacy Notice needs updating.
Getting schooled on your own domain expertise by an AI is... a peculiar feeling.
The Real Takeaway
Today's biggest takeaway wasn't installing 13 agents — it was watching two personas analyze the same system with fundamentally different mental models. This is something plain prompt engineering can't deliver. You can tell an AI to "act as a security expert," but you can't give it the instinct of a seasoned security professional who spots the real risk at first glance. That instinct is written into the persona, not the prompt.
The RPG Era
The RPG era is truly here. It used to take three visits to the thatched cottage to summon a Zhuge Liang. Now all it takes is a markdown file.
But this also means — the ceiling of what you can do with AI no longer depends on whether you can write prompts. It depends on whether you can read people, delegate work, and evaluate reports. That's a management skill, not a technical skill.
Do the work of Zhuge Liang. Live the life of Su Shi.
Only now, when you wave the feathered fan, the troops you deploy are markdown files.