mirror of
https://github.com/hanzoai/papers.git
synced 2026-10-29 07:34:31 +00:00
zeekay
0d1d121451
Fix shared/ include paths to ../shared/ for 55 subdirectory papers so
that hanzocover.sty and lstlang.tex resolve when building from each
paper's own directory (the build pattern used by latexmk per paper).
Build fixes:
- hanzo-operative/sections/03-screen-understanding-pipeline.tex: move
\COMMENT from after \FOR{...} onto the \STATE line so the algorithmic
env doesn't mis-parse.
- hanzo-network-architecture (abstract + sections 04, 12, 13): escape
unicode mu (U+03BC) as $\mu$ for pdflatex.
- ntt-transform/sections/08-benchmark-results.tex: same mu escape.
All 55 Tier-2/3 papers now build clean with latexmk -pdf.
67 lines
3.0 KiB
TeX
67 lines
3.0 KiB
TeX
% Hanzo Flow: Visual Workflow Builder for AI Agent Orchestration
|
|
\documentclass[11pt,twocolumn]{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{float}
|
|
\usepackage{subcaption}
|
|
\usepackage{multirow}
|
|
\usepackage{listings}
|
|
\input{../shared/lstlang}
|
|
\usepackage[numbers]{natbib}
|
|
\hypersetup{colorlinks=true,linkcolor=black,citecolor=blue,urlcolor=blue}
|
|
|
|
\newtheorem{theorem}{Theorem}
|
|
\newtheorem{lemma}[theorem]{Lemma}
|
|
\newtheorem{definition}{Definition}
|
|
\newtheorem{proposition}[theorem]{Proposition}
|
|
\newtheorem{corollary}[theorem]{Corollary}
|
|
|
|
\lstset{
|
|
basicstyle=\ttfamily\small,
|
|
breaklines=true,
|
|
frame=single,
|
|
numbers=left,
|
|
numberstyle=\tiny,
|
|
xleftmargin=2em,
|
|
}
|
|
|
|
\title{Hanzo Flow: Visual Workflow Builder for AI Agent Orchestration}
|
|
\author{
|
|
David Wei \quad
|
|
Marcus Chen \quad
|
|
Zach Kelling \\
|
|
\textit{Hanzo AI Research} \\
|
|
\texttt{[email protected]}
|
|
}
|
|
\date{April 2025}
|
|
|
|
\begin{document}
|
|
|
|
\hanzocoverpage
|
|
|
|
\begin{abstract}
|
|
We present \textbf{Hanzo Flow}, a visual workflow builder for constructing, debugging, and deploying multi-agent AI systems. While code-based agent frameworks provide flexibility, they impose high barriers to iteration, lack visibility into agent decision-making, and make debugging multi-agent interactions intractable. Hanzo Flow introduces three innovations: (i) a \emph{typed dataflow graph} formalism where nodes represent agent operations (LLM calls, tool invocations, control flow, human-in-the-loop gates) and edges represent typed data channels, enabling static verification of workflow correctness before execution; (ii) a \emph{visual debugger} with execution replay, token-level attribution, and branching exploration that reduces debugging time by 67\% compared to log-based debugging; and (iii) an \emph{adaptive execution engine} that supports parallel branch execution, dynamic subgraph expansion, and automatic retry with model fallback, achieving 94\% end-to-end success rate on complex multi-step tasks. We evaluate Hanzo Flow on the AgentBench suite, a novel visual workflow benchmark (FlowBench-200), and a user study with 48 developers. Results show that visual workflow construction reduces development time by 3.2x compared to code-based frameworks while achieving equivalent or superior task completion rates. Hanzo Flow has been used in production to build 1,200+ workflows processing 8.4 million agent invocations over 10 months.
|
|
\end{abstract}
|
|
|
|
\input{sections/01-introduction}
|
|
\input{sections/02-typed-dataflow-graphs}
|
|
\input{sections/03-visual-debugger}
|
|
\input{sections/04-adaptive-execution-engine}
|
|
\input{sections/05-workflow-patterns}
|
|
\input{sections/06-evaluation}
|
|
\input{sections/07-production-deployment}
|
|
\input{sections/08-related-work}
|
|
\input{sections/09-discussion}
|
|
\input{sections/99-conclusion}
|
|
|
|
\end{document}
|