Author: ParisNeo
Pagination Engine: LoLLMs (Lord of Large Language and Multimodal Systems)
Version: 1.0.0
Date: May 06, 2026
PREAMBLE: The 98.4% Truth
An agentic harness is the 98.4% of your system that isn’t the model—it is the deterministic infrastructure (permission gates, context management, and recovery logic) that makes an agent viable.
This manifesto outlines how to build an industry-leading harness while avoiding the ethical and security failures revealed in the Claude Code leak.
“The model is the soul, but the harness is the skeleton. Without a strong skeleton, the soul cannot act.”
— ParisNeo
I. THE EFFICIENCY PILLARS: Mastering Context & Memory
The greatest constraint on agents is the context window. Use a “graduated” approach rather than a single-pass truncation.
1.1 The Three-Layer Memory System
| Layer | Name | Purpose | Max Size |
|---|---|---|---|
| Layer 1 | Index (memory.md) | Table of contents for your project | 200 lines |
| Layer 2 | Topic Files | Fetch specific project details on demand | Variable |
| Layer 3 | Transcripts | Pointer-based retrieval for session history | Never full context |
Implementation Rule: Always fetch on demand. Never load the whole project into context.
1.2 Five-Layer Compaction Pipeline
Run these sequentially (cheapest first) before every model call: