Skip to content

v1.0 — Initial Publication

Choose a tag to compare

View all tags
@fenix163 fenix163 released this 28 May 23:19
v1.0
b28e682

Context Isolation in Multi-Agent LLM Systems

First public release of the research paper and experimental data.

Paper: main.pdf (18 pages, 456KB)
Data: 23 experimental runs across 3 execution paradigms
Code: score_run.py, run_experiments.py

Key Finding

The hypothesis that context isolation through subagent decomposition uniformly improves task quality is not supported. Single-session TUI outperforms vanilla Kanban on tightly-coupled multi-file tasks (Cohen's d = 1.24).

Abstract

We compare three execution paradigms in the Hermes Agent framework: single-session TUI, vanilla Kanban orchestrator with generic decomposition, and Self-Tuned Kanban (STK) with explicit interface contracts. Using a novel Context Isolation Rubric (CIR) across five dimensions on 23 experimental runs, we evaluate code refactoring, research synthesis, and multi-file bug diagnosis challenges. Results indicate that the optimal paradigm depends on task structure — more agents is not always better.

Author: David Scott, NovaFuse Inc., Ottawa, Canada
Email: info@novafuse.ca

Assets 3
Loading