papers/morpheus/onepagers/lp-317-session-agents.tex

\documentclass[10pt,a4paper]{article}
\usepackage{../shared/proposal}
\usepackage[utf8]{inputenc}\usepackage[T1]{fontenc}
\usepackage{amsmath,amssymb}\usepackage{listings}\usepackage{xcolor}
\input{../shared/lstlang}

\begin{document}
\sloppy

\proposalhead{Sessions as the Agent Substrate}{Per-second billing is a poor fit for chat, where the
session is mostly silence. It is close to the right fit for an agent, whose idle time is the
work.}{LP-317}

\section*{The problem}

An agent turn is long-lived, serial and tool-heavy, and its gaps are the agent working rather than
a human pausing, so it is the one workload a per-second clock does not overcharge. Morpheus already
bills in that unit and carries tool calls over the session, and holds no trajectory across either.

\section*{The mechanism}

A trajectory alternates phases $(g_1,t_1,\dots,g_n,t_n)$: $g_i$ seconds generating at
$\bar r$ tokens per second, $t_i$ seconds blocked on a tool, $G=\sum g_i$, $T=G+\sum t_i$,
$\rho=G/T$. A per-token price $\kappa$ charges $\kappa\bar rG$ and ignores every $t_i$; a
per-second bid $p$ charges $pT$, an effective $p/(\bar r\rho)$ per token. They agree at
$\rho^{\star}=p/(\kappa\bar r)$: chat sits far below it, an agent burst against local tools near
$\rho=1$, and per-second is worse exactly where $\rho$ collapses, forty seconds on a slow external
tool buying forty seconds of GPU price and no work.

It is an admission problem before a pricing one, and admission is factored well already.
\texttt{HasCapacity(modelID)} is the whole interface (\texttt{capacity\_manager.go}:12):
\texttt{simple} counts open sessions, \texttt{idle\_timeout} ones active inside a hardcoded fifteen
minutes (:73), the right instinct at the wrong resolution. \texttt{PromptActivity} carries \texttt{SessionID,
StartTime, EndTime} (\texttt{structs.go}:32), enough to seat a second consumer in the slot a parked
one leaves idle. One more case in \texttt{CreateCapacityManager} (:112).

\section*{What a trajectory needs that a session is not}

\textbf{Tools run on the provider.} \texttt{CallAgentTool(ctx, sessionID, agentID, \dots)}
(\texttt{ai\_engine.go}:195) runs a registered agent's \texttt{command}
(\texttt{agents\_config.go}:32) as an MCP stdio child of the provider process (:224). The provider
executes the tool on its own credentials, paid per second for as long as it takes, and the consumer
authorises by holding a session without seeing what ran. A provider running slow tools is paid
more: admission and settlement are one question.

\textbf{A session spans neither a second bid nor its own funding.} History keys on
\texttt{chat\_id}, a client header randomised when absent (\texttt{controller\_http.go}:217); tools
and payment key on $\texttt{sessionID}=H(\text{user},\text{provider},\text{bidId},\text{nonce})$
(\texttt{SessionRouter.sol}:159), so a trajectory wanting two models holds two sessions whose tool
state cannot join although their \texttt{chat\_id} does. And \texttt{endsAt} is fixed at open by
$\text{stipend}/p$ (:168) with no extend, so outrunning the funding resumes the task under a new
\texttt{sessionId}, history intact, tool handles gone.

One client-side object closes both: a \texttt{chat\_id} and a set of live \texttt{sessionId}s,
opened before the current expires rather than after, one per bid, presenting one tool namespace
across them. Latency under a per-second bid \emph{is} cost, so the policy falls out: hold the cheap
session open across waits, spend into the expensive one only while generating.

\begin{stake}{Morpheus gets}
\item Admission on generating load, more consumers per GPU at an unchanged bid, and the first
      measured $\rho$, a figure absent from all 62 repositories.
\item An agent story on the marketplace. Today it is seven forks: five Vercel-derived templates set
      an OpenAI base URL to \texttt{api.mor.org/api/v1} and never touch a session, while
      \texttt{sim-studio-morpheus} and \texttt{Perplexica-morpheus} carry the name and nothing
      else.
\end{stake}

\begin{stake}{Hanzo and Lux get}
\item A settled, disputable market in the unit agents consume, and a live price for Enso's cost
      multiplier.
\end{stake}

\section*{First step, and who takes it}

Hanzo writes the third capacity manager against \texttt{dev} and publishes $\rho$, generate against
paid seconds, over its own provider's traffic. No facet upgrade, no 5-of-9 signature, and the
figure is the network's either way.

\section*{What this does not solve}

A trajectory makes provider-side tool execution cheaper to buy and no easier to verify: nothing
attests that the tool ran, as nothing attests that the advertised model answered.

\vspace{0.5ex}
\noindent{\color{muted}\footnotesize\sffamily LP-317, full paper available. Market diamond \texttt{0x6aBE1d282f72B474E54527D93b979A4f64d3030a} on Base. See LP-315, LP-313.}

\end{document}