Thin Soul, Fat Skill
The architecture principle that lets 74 agents fit in 960MB on a Mac Mini.
The Problem
Most agent frameworks embed business logic inside prompts. This makes prompts bloat to thousands of tokens, wastes inference on deterministic operations, and couples identity to execution. When you want to change behavior, you rewrite the prompt. When you want to change identity, you also rewrite the prompt.
The Principle
Separate identity from execution logic:
- ●Soul — a lightweight YAML+Markdown file (~30 lines) that declares who the agent is, what skills it can use, what safety rules apply. This becomes the system prompt.
- ●Skills — deterministic plugins (Python/bash/Go) that handle data retrieval, computation, API calls. No LLM involvement.
Why It Works
Token efficiency: LLM only sees declarative config + conversation. Deterministic ops (file I/O, HTTP, math) cost zero tokens.
Hot swapping: Change the soul file, agent personality changes. No rebuild.
Machine modifiable: swarm_architect can edit soul files autonomously because they're text.
Composable: Same skill used by 40 agents. No duplication.
Example
A stock price lookup takes 50 tokens through the Fat Skill pattern:
stock_price action="quote" symbol="NVDA" → {"price": 177.39, ...}
The same operation via prompt engineering would cost thousands of tokens and risk hallucination.
Related
- ●The 74-Agent Swarm — how this scales
- ●The Heart System — how souls stay alive
- ●Knowledge Layers — where source-of-truth lives