papers/hanzo-unified-harness/hanzo-unified-harness.tex

% The Unified Agent Harness: A Vertically Integrated Architecture for Autonomous AI Agents
% Fully self-contained -- no \input{} directives
\documentclass[11pt]{article}
\usepackage{../shared/hanzocover}
\usepackage{amsmath, amssymb, amsthm}
\usepackage{mathtools}
\usepackage{bm}
\usepackage{graphicx}
\usepackage{booktabs}
\usepackage{hyperref}
\usepackage{enumitem}
\usepackage{algorithm}
\usepackage{algpseudocode}
\usepackage{xcolor}
\usepackage[numbers]{natbib}
\usepackage{float}
\usepackage{subcaption}
\usepackage{tabularx}
\usepackage{multirow}

\hypersetup{colorlinks=true,linkcolor=black,citecolor=blue,urlcolor=blue}

\newtheorem{theorem}{Theorem}
\newtheorem{lemma}[theorem]{Lemma}
\newtheorem{proposition}[theorem]{Proposition}
\newtheorem{corollary}[theorem]{Corollary}
\newtheorem{definition}{Definition}
\newtheorem{remark}{Remark}

\DeclareMathOperator*{\argmin}{arg\,min}
\DeclareMathOperator*{\argmax}{arg\,max}
\DeclareMathOperator{\KL}{KL}
\DeclareMathOperator{\Tr}{Tr}
\DeclareMathOperator{\diag}{diag}

\definecolor{hanzoviolet}{rgb}{0.42,0.18,0.82}
\definecolor{hanzoteal}{rgb}{0.07,0.55,0.55}
\definecolor{hanzogray}{rgb}{0.45,0.45,0.45}

\title{The Unified Agent Harness: A Vertically Integrated\\
       Architecture for Autonomous AI Agents}
\author{
    Hanzo AI Research \\
    Hanzo AI Research\\
    \textit{Hanzo AI Inc (Techstars '17), San Francisco, CA}\\
    \texttt{[email protected]}
}
\date{February 2026}

\begin{document}

\hanzocoverpage

% ============================================================================
% ABSTRACT
% ============================================================================
\begin{abstract}
We present the \textbf{Unified Agent Harness (UAH)}, a vertically integrated
runtime architecture for autonomous AI agents that resolves the fundamental
mismatch between the stochastic, open-ended nature of language model inference
and the deterministic, bounded interfaces of existing software toolchains.
Current agent platforms---including Claude Code~\cite{anthropic2024claude},
Cursor~\cite{cursor2024}, Devin~\cite{cognition2024devin}, and OpenAI
Codex~\cite{openai2023gpt4}---treat the agent as a \emph{feature} of an
existing product category (IDE, chat interface, or CI pipeline) rather than as
a first-class computational primitive. We argue this framing creates a
structural ceiling on agent capability, analogous to running an operating system
as a process within another operating system.

The UAH defines a seven-layer architecture spanning LLM access, tool protocol,
agent runtime, channel integration, compute tiers, self-improvement, and
observability, and argues that cross-layer optimization across all seven---only
achievable through vertical integration---is necessary and sufficient for
production-grade autonomous agents. We formalize each layer as a typed interface
with explicit invariants, define a five-tier compute hierarchy with formal
state-machine transitions, and prove that cross-layer optimization strictly
dominates any composition of independently optimized layers. Empirical
evaluation across four agent deployment domains (software engineering, marketing
automation, DevOps operations, and scientific research) demonstrates that the
UAH achieves mean task-completion rates of \textbf{74.3\%} on SWE-bench
Verified~\cite{jimenez2024swebench}, compared to 43.7\% for the best
single-vendor competitor, with $2.1\times$ lower median cost per resolved task.

The central thesis of this paper is architectural: the \emph{harness}, not the
model, is the product. As frontier model capabilities converge, the agent
infrastructure layer becomes the durable source of competitive differentiation.
\end{abstract}

% ============================================================================
% 1. INTRODUCTION
% ============================================================================

\input{sections/01-introduction}
\input{sections/02-related-work}
\input{sections/03-the-seven-layer-architecture}
\input{sections/04-tiered-compute-runtime}
\input{sections/05-cross-layer-optimization}
\input{sections/06-competitive-analysis}
\input{sections/07-multi-domain-generalization}
\input{sections/08-production-implementation}
\input{sections/09-discussion-and-future-work}
\input{sections/10-conclusion}
\input{sections/11-complete-layer-interface-specifications}
\input{sections/12-swe-bench-evaluation-setup}
\input{sections/99-tier-complexity-thresholds}

\section*{Reproducibility}
\textbf{Source code.} \texttt{github.com/hanzoai/agents} (commit \texttt{TBD}); seven-layer reference impl across \texttt{hanzoai/engine}, \texttt{hanzoai/mcp}, \texttt{hanzoai/agents}, \texttt{hanzoai/base}.\\
\textbf{Build.} \texttt{make all} at repo root; integration via \texttt{kubectl kustomize universe/infra/k8s/ | kubectl apply -f -}.\\
\textbf{Hardware tested.} Apple M-series (Metal), Linux amd64/arm64 with NVIDIA A100/H100; tiered runtime validated on M2 Pro and H100.\\
\textbf{Standards referenced.} MCP, OpenAI tool-use API, SWE-bench evaluation protocol.\\
\textbf{Contact.} [email protected].

\end{document}