mirror of
https://github.com/hanzoai/papers.git
synced 2026-11-03 02:27:30 +00:00
zeekay
0f01f8b555
Resolves high-severity review findings: anachronistic dates aligned to citation floors; arithmetic corrected to match formulas; internal contradictions collapsed to the body-supported value; LP self-references aligned to filenames; liquidity-surface repo paths scrubbed; AI-slop trimmed. PDFs rebuilt separately.
77 lines
3.3 KiB
TeX
77 lines
3.3 KiB
TeX
\documentclass[11pt,twocolumn]{article}
|
|
\usepackage{../shared/hanzocover}
|
|
\usepackage[utf8]{inputenc}
|
|
\usepackage{amsmath,amssymb,amsthm}
|
|
\usepackage{graphicx}
|
|
\usepackage{booktabs}
|
|
\usepackage{algorithm}
|
|
\usepackage{algorithmic}
|
|
\usepackage{hyperref}
|
|
\usepackage{xcolor}
|
|
\usepackage{listings}
|
|
\input{../shared/lstlang}
|
|
\usepackage{tikz}
|
|
\usetikzlibrary{shapes,arrows,positioning,fit,backgrounds,calc}
|
|
\usepackage{enumitem}
|
|
\usepackage{tabularx}
|
|
\usepackage{multirow}
|
|
\usepackage{subcaption}
|
|
|
|
\definecolor{codegreen}{rgb}{0,0.6,0}
|
|
\definecolor{codegray}{rgb}{0.5,0.5,0.5}
|
|
\definecolor{codepurple}{rgb}{0.58,0,0.82}
|
|
\definecolor{backcolour}{rgb}{0.95,0.95,0.92}
|
|
|
|
\lstdefinestyle{mystyle}{
|
|
backgroundcolor=\color{backcolour},
|
|
commentstyle=\color{codegreen},
|
|
keywordstyle=\color{codepurple},
|
|
numberstyle=\tiny\color{codegray},
|
|
stringstyle=\color{codegreen},
|
|
basicstyle=\ttfamily\footnotesize,
|
|
breakatwhitespace=false,
|
|
breaklines=true,
|
|
captionpos=b,
|
|
keepspaces=true,
|
|
numbers=left,
|
|
numbersep=5pt,
|
|
showspaces=false,
|
|
showstringspaces=false,
|
|
showtabs=false,
|
|
tabsize=2
|
|
}
|
|
\lstset{style=mystyle}
|
|
|
|
\newtheorem{definition}{Definition}
|
|
\newtheorem{theorem}{Theorem}
|
|
\newtheorem{lemma}{Lemma}
|
|
\newtheorem{proposition}{Proposition}
|
|
|
|
\title{Operative: A Computer Use Framework\\for Autonomous AI Agents}
|
|
\author{Antje Worring, Zach Kelling \\
|
|
Hanzo AI Research\\
|
|
\texttt{[email protected]}\\
|
|
\textit{Hanzo AI \quad San Francisco, CA}}
|
|
\date{2026}
|
|
|
|
\begin{document}
|
|
|
|
\hanzocoverpage
|
|
|
|
\begin{abstract}
|
|
Autonomous AI agents capable of operating graphical user interfaces represent a significant frontier in applied artificial intelligence. We present Operative, an open-source framework for computer use that enables large language models to perceive, reason about, and interact with desktop and browser environments. Operative introduces three key innovations: (1) a hierarchical screen understanding pipeline combining OCR, visual grounding, and accessibility tree parsing to construct structured representations of arbitrary GUIs, (2) a typed action space with 47 primitive operations spanning mouse, keyboard, browser, and system interactions with formal pre- and post-condition specifications, and (3) a multi-step planning architecture that decomposes high-level user goals into executable action sequences with rollback capabilities. Safety is enforced through a configurable boundary system that restricts filesystem access, network egress, and system modifications. In evaluation across three established benchmarks---OSWorld, WebArena, and VisualWebBench---Operative achieves task completion rates of 38.1\%, 54.3\%, and 52.8\% respectively with a Claude backend, establishing state-of-the-art results across all three. The framework processes 4.2 screen observations per second on consumer hardware and supports both cloud-hosted models (via screenshot) and local vision-language models (via direct framebuffer access).
|
|
\end{abstract}
|
|
|
|
\input{sections/01-introduction}
|
|
\input{sections/02-background-and-related-work}
|
|
\input{sections/03-screen-understanding-pipeline}
|
|
\input{sections/04-action-space}
|
|
\input{sections/05-safety-boundaries}
|
|
\input{sections/06-multi-step-planning}
|
|
\input{sections/07-implementation}
|
|
\input{sections/08-evaluation}
|
|
\input{sections/09-discussion}
|
|
\input{sections/99-conclusion}
|
|
|
|
\end{document}
|