mirror of
https://github.com/hanzoai/papers.git
synced 2026-10-30 14:58:42 +00:00
116 lines
4.9 KiB
TeX
116 lines
4.9 KiB
TeX
\section{Multi-Domain Generalization}
|
|
|
|
\label{sec:domains}
|
|
|
|
A key architectural claim of the UAH is that the seven-layer harness is
|
|
\textbf{domain-agnostic}: the same infrastructure serves agents across
|
|
radically different task types, with domain specificity concentrated in the
|
|
Layer 3 agent specifications (skills, tools, instructions) rather than in
|
|
the infrastructure layers.
|
|
|
|
We formalize this claim:
|
|
|
|
\begin{definition}[Domain-Agnostic Harness]
|
|
A harness $H$ is \emph{domain-agnostic} if for any two task domains
|
|
$\mathcal{D}_1$ and $\mathcal{D}_2$, the infrastructure layers $L_1, L_2,
|
|
L_4, L_5, L_6, L_7$ are identical and only $L_3$ (Agent Runtime)
|
|
configurations differ.
|
|
\end{definition}
|
|
|
|
We demonstrate this property across four deployment domains.
|
|
|
|
\subsection{Software Engineering Domain}
|
|
|
|
The software engineering agent (``Hanzo Dev'') uses:
|
|
\begin{itemize}[leftmargin=1.5em]
|
|
\item \textbf{L3 configuration:} Orchestrator agent + specialist agents
|
|
for planning, code generation, testing, and review.
|
|
\item \textbf{Primary tools:} Code intelligence (10 tools), shell
|
|
execution (8 tools), filesystem (12 tools).
|
|
\item \textbf{Default tier:} $T_1$ (Terminal Pod). Upgrades to $T_3$ for
|
|
CI/CD tasks requiring cloud services.
|
|
\item \textbf{Improvement signal:} Test pass rate, linting score, PR merge
|
|
acceptance rate.
|
|
\end{itemize}
|
|
|
|
Performance on SWE-bench Verified: 74.3\% resolved rate.
|
|
|
|
\subsection{Marketing Automation Domain}
|
|
|
|
The marketing agent (``Hanzo Growth'') uses:
|
|
\begin{itemize}[leftmargin=1.5em]
|
|
\item \textbf{L3 configuration:} Research agent + copywriter agent +
|
|
analytics agent + publisher agent.
|
|
\item \textbf{Primary tools:} Web search, browser automation, data
|
|
analysis, email/CMS publishing tools (via custom MCP extensions).
|
|
\item \textbf{Default tier:} $T_0$ (Shared Gateway) for research and
|
|
copywriting. $T_2$ (Operative Desktop) for browser-based campaign
|
|
management.
|
|
\item \textbf{Improvement signal:} Click-through rate, conversion rate,
|
|
campaign ROI.
|
|
\end{itemize}
|
|
|
|
The marketing domain illustrates the Channel Integration layer's value:
|
|
campaigns are triggered from Slack, executed autonomously, and results are
|
|
reported back to Slack and the connected Linear project.
|
|
|
|
\subsection{DevOps Operations Domain}
|
|
|
|
The DevOps agent (``Hanzo Ops'') uses:
|
|
\begin{itemize}[leftmargin=1.5em]
|
|
\item \textbf{L3 configuration:} Alert triage agent + diagnosis agent +
|
|
remediation agent, connected to PagerDuty webhook via Layer 4.
|
|
\item \textbf{Primary tools:} kubectl, Helm, Prometheus queries, log
|
|
analysis, GitHub (deploy workflows).
|
|
\item \textbf{Default tier:} $T_3$ (Cloud VM with direct cluster access).
|
|
\item \textbf{Improvement signal:} Mean time to resolution (MTTR),
|
|
false positive alert rate, successful remediation rate.
|
|
\end{itemize}
|
|
|
|
Over 90 days of production deployment, Hanzo Ops reduced mean MTTR from 47
|
|
minutes to 8.3 minutes for the class of incidents within its capability scope
|
|
(database performance issues, deployment failures, certificate expirations,
|
|
and capacity alerts).
|
|
|
|
\subsection{Research Domain}
|
|
|
|
The research agent (``Hanzo Research'') uses:
|
|
\begin{itemize}[leftmargin=1.5em]
|
|
\item \textbf{L3 configuration:} Literature search agent + experimental
|
|
design agent + data analysis agent + writing agent.
|
|
\item \textbf{Primary tools:} ArXiv/PubMed search, PDF parsing, statistical
|
|
analysis tools, Python/R execution, citation management.
|
|
\item \textbf{Default tier:} $T_3$ (Cloud VM for compute-heavy analysis).
|
|
Upgrades to $T_4$ (Local Desktop) when GPU acceleration is required.
|
|
\item \textbf{Improvement signal:} Literature coverage completeness,
|
|
statistical validity (verified by Reviewer agent), citation accuracy.
|
|
\end{itemize}
|
|
|
|
\begin{table}[H]
|
|
\centering
|
|
\caption{Domain-Specific Performance Metrics}
|
|
\label{tab:domains}
|
|
\begin{tabular}{lllcc}
|
|
\toprule
|
|
\textbf{Domain} & \textbf{Primary Metric} & \textbf{UAH Score} &
|
|
\textbf{Default Tier} & \textbf{Improvement Loops} \\
|
|
\midrule
|
|
Software Eng. & SWE-bench resolved rate & 74.3\% & $T_1$ & 4 \\
|
|
Marketing Auto. & Campaign setup accuracy & 81.2\% & $T_0$/$T_2$ & 3 \\
|
|
DevOps Ops. & Incident resolution rate & 76.8\% & $T_3$ & 4 \\
|
|
Research & Literature coverage & 88.4\% & $T_3$/$T_4$ & 4 \\
|
|
\midrule
|
|
\multicolumn{2}{l}{\textbf{Mean}} & \textbf{80.2\%} & --- & --- \\
|
|
\bottomrule
|
|
\end{tabular}
|
|
\end{table}
|
|
|
|
Across all four domains, the infrastructure layers (L1, L2, L4, L5, L6, L7)
|
|
are identical. Total additional code required to deploy a new domain is
|
|
approximately 400--800 lines: the Layer 3 agent specification, any custom
|
|
MCP tool extensions, and the domain-specific improvement signal definition.
|
|
|
|
% ============================================================================
|
|
% 8. PRODUCTION IMPLEMENTATION
|
|
% ============================================================================
|