The Counterintuitive Truth
When people hear "224 AI employees," they assume we wrote 224 separate prompts. We didn't. We wrote exactly 4 core templates. Everything else is structured config data.
This is the full postmortem of our journey from prompt explosion (180,000 words, 2 hours per new role) to a templated system (2,400 words, 5 minutes per new role).
Why We Needed 224 Roles
Cross-border trade is complex. Content, marketing, sales, logistics, compliance, finance, legal — each requires different expertise. We tried a single "super-agent." It produced sloppy legal reviews, stiff marketing copy, and leaked financial data into public-facing content.
So we modeled a real trade company: 18 departments, 224 clearly defined roles.
The Prompt Explosion Nightmare
In v1, we wrote a unique prompt for every role. ~800 words each. ~180,000 words total. Three people, two full weeks.
Then the nightmare:
- Changing one rule meant editing 224 files
- Roles drifted and blended into each other
- New role onboarding took ~2 hours
- Different departments ran different rule versions
We even tried hiring interns to maintain prompts. Dead end.
The Solution: Three-Layer Templated Architecture
After four iterations, we landed on: Template → Role Card → Federated Coordination
Layer 1: Core Template (1 universal)
12 fixed fields: identity, outcomes, boundaries, output format, communication protocol, compliance guardrails. Every role inherits this skeleton.
Layer 2: Department Constraints (18 sets)
Department-specific rules injected once, applied to all roles in that team. Finance, Legal, Marketing — each has its own constraint layer.
Layer 3: Role Parameters (224 sets, auto-injected)
Only 5–8 parameters per role, stored as YAML/JSON config files. No free-text prompts.
Global Layer: Federated Coordination Rules
Task routing, handoff protocol, conflict resolution, audit requirements — one set for the entire organization.
Before vs. After
| Metric | v1 (224 prompts) | v2 (templates) | Improvement |
|---|---|---|---|
| Prompt volume | ~180,000 words | ~2,400 words | 98.7% |
| New role creation | ~2 hours | ~5 minutes | 24x faster |
| Global rule update | ~8 hours | ~10 minutes | 48x faster |
| Role compliance rate | 72% | 94% | +22 pp |
| Handoff failure rate | 31% | 8% | 74% ↓ |
How It Works in Practice
When a user says "launch a new skincare product in Germany":
- System decomposes into subtasks (regulatory, marketing, logistics, pricing)
- Routes each to the correct department and role
- Runs independently with role-specific constraints
- Orchestrates cross-department handoffs
- Aggregates into a single final deliverable
No single agent holds full context — just like a real company.
Join the Discussion
The full discussion is on GitHub: How 224 AI Employees "Apply" for Their Jobs
We'd love to hear how you handle role management and cross-agent coordination in your multi-agent systems.
Next in the series: How we run performance reviews for 224 AI employees — KPIs, automated evaluation loops, and dynamic role optimization.
Top comments (0)