papers/hanzo-dormant-agents/hanzo-dormant-agents.tex

% What a Dormant Agent Costs
\documentclass[11pt]{article}
\usepackage{../shared/hanzocover}
\usepackage{amsmath, amssymb}
\usepackage{booktabs}
\usepackage{graphicx}
\usepackage{xcolor}
\usepackage{enumitem}
\usepackage{array}
\usepackage{listings}
\input{../shared/lstlang}
\usepackage{hyperref}
\hypersetup{colorlinks=true,linkcolor=black,citecolor=blue,urlcolor=blue}

% Provenance, on every number. A reader should never have to guess whether a
% figure came off a machine or out of a spreadsheet.
\newcommand{\meas}{\textcolor{blue!60!black}{\small[measured]}}
\newcommand{\modeled}{\textcolor{orange!70!black}{\small[modeled]}}
\newcommand{\cited}{\textcolor{purple!70!black}{\small[cited]}}

% Every measured constant is a macro, and results-data.tex overrides all of
% them from the runs. The defaults below are placeholders: if the generated
% file is ever missing, the paper prints XXX rather than a stale number that
% reads as real.
\newcommand{\ph}{\textbf{XXX}}
\newcommand{\rFleetAgents}{\ph}   \newcommand{\rFleetRecord}{\ph}
\newcommand{\rFleetDisk}{\ph}     \newcommand{\rFleetPer}{\ph}
\newcommand{\rFleetWriteS}{\ph}   \newcommand{\rFleetRate}{\ph}
\newcommand{\rFleetResume}{\ph}   \newcommand{\rFleetResumeCold}{\ph}
\newcommand{\rGoSpawn}{\ph}       \newcommand{\rGoSpawnLo}{\ph}
\newcommand{\rGoSpawnHi}{\ph}     \newcommand{\rGoPer}{\ph}
\newcommand{\rGoHeap}{\ph}        \newcommand{\rGoWake}{\ph}
\newcommand{\rGoWakeLo}{\ph}      \newcommand{\rGoWakeHi}{\ph}
\newcommand{\rWazero}{\ph}        \newcommand{\rWazeroHi}{\ph}
\newcommand{\rWazeroCall}{\ph}    \newcommand{\rGoja}{\ph}
\newcommand{\rGojaLo}{\ph}        \newcommand{\rGojaHi}{\ph}
\newcommand{\rGojaWarm}{\ph}      \newcommand{\rGojaWarmLo}{\ph}
\newcommand{\rGojaWarmHi}{\ph}    \newcommand{\rGpython}{\ph}
\newcommand{\rGpythonLo}{\ph}     \newcommand{\rGpythonHi}{\ph}
\newcommand{\rVeight}{\ph}        \newcommand{\rVeightP}{\ph}
\newcommand{\rVm}{\ph}            \newcommand{\rVmP}{\ph}
\newcommand{\rVmCkpt}{\ph}        \newcommand{\rVmCkptP}{\ph}
\newcommand{\rVmRatio}{\ph}       \newcommand{\rPlaneCost}{\ph}
\newcommand{\rPlaneCpu}{\ph}      \newcommand{\rPlaneRam}{\ph}
\newcommand{\rFleetStorage}{\ph}
\input{results-data}

\title{What a Dormant Agent Costs\\[0.3em]
{\large State, Execution and Isolation Measured Separately\\
at a Fleet of One Million}}
\author{
    Hanzo AI Research\thanks{Corresponding author: [email protected]} \\
    \textit{Hanzo Industries Inc (Techstars '17)} \\
    San Francisco, California \\
    \texttt{https://hanzo.ai}
}
\date{Measurement pass: 2026-09-07 (Apple M-series, darwin/arm64, 16 cores, Go 1.27.1)}

\begin{document}

\hanzocoverpage
\maketitle

\begin{abstract}
An agent platform is usually costed as though an agent were a process. It is the
natural reading of the deployment shape --- one container per agent --- and it
sets the per-agent floor at the container runtime's minimum reservation, which
on commodity platforms is 128--512\,MB\@. At that floor a fleet of a million
agents is a datacenter, and every design decision downstream of the estimate is
made in its shadow.

We take the estimate apart. An idle agent is three separable things: the
\emph{state} it will resume from, the \emph{execution context} that holds its
loop, and the \emph{isolation boundary} its code runs behind. These have no
reason to cost the same, and measured on one laptop they differ by five orders
of magnitude. State at rest is a row: \textbf{\rFleetPer{}\,bytes} on disk
\meas{}, so \rFleetAgents{} dormant agents occupy \textbf{\rFleetDisk{}\,MiB}
--- written at \rFleetRate{}\,agents/s and resumed in
\textbf{\rFleetResume{}\,ms} in-process. Execution at rest is a goroutine:
\textbf{\rGoPer{}\,bytes} of heap \meas{}, identical on every run, so a million
live loops are \rGoHeap{}\,MB and each starts in \rGoSpawn{}\,ns. Isolation is
the expensive one, and it is expensive for a reason: an in-process V8 context
costs \rVeight{}\,ms \meas{} but runs only JavaScript, while a microVM
costs \rVm{}\,ms --- \rVmRatio{}$\times$ more --- and is the only one of the
two that can run a test suite, a package manager or a shell.

The conclusion is not that containers are wasteful. It is that a fleet pays for
isolation \emph{per running agent}, not per registered one, while the industry
estimate charges the isolation price to every agent in the fleet whether or not
it is doing anything. Separating the three costs makes the dormant majority
nearly free: \rFleetAgents{} dormant agents are \rFleetStorage{} dollars a month
of storage \modeled{} beside a control plane that costs \$\rPlaneCost{}
\cited{}. We give the capacity model this implies, and we are explicit in
\S\ref{sec:limits} about what \rGoPer{}\,bytes is: a floor for an agent holding
nothing, not a forecast for an agent holding a conversation.
\end{abstract}

\input{sections/01-introduction}
\input{sections/02-method}
\input{sections/03-state}
\input{sections/04-execution}
\input{sections/05-code}
\input{sections/06-isolation}
\input{sections/07-capacity}
\input{sections/08-limits}
\input{sections/99-conclusion}

\end{document}