papers/lux-agent-sovereignty/sections/01-introduction.tex

\section{Introduction}
\label{sec:intro}

The current generation of AI agent frameworks treats agents as ephemeral function calls. An agent is instantiated, executes a sequence of API requests, and terminates. It holds no persistent state, owns no cryptographic identity, controls no assets, and produces no verifiable record of its actions. This model is adequate for single-turn chatbots. It is wholly inadequate for autonomous agents that must operate in adversarial, multi-party, economically consequential environments.

Consider the requirements of a production AI agent:

\begin{enumerate}[label=(\roman*)]
    \item \textbf{Persistence.} The agent must maintain state across sessions---learned preferences, accumulated context, intermediate results of long-running tasks.
    \item \textbf{Privacy.} The agent's reasoning trace, memory, and intermediate computations must remain confidential, even from the infrastructure operator.
    \item \textbf{Accountability.} Every action the agent takes must be attributable, auditable, and non-repudiable. If the agent signs a transaction, there must be a verifiable chain from action to identity.
    \item \textbf{Composability.} Agents must interact with other agents, delegate subtasks, share partial results, and coordinate on joint objectives---all without trusting each other.
    \item \textbf{Economic rationality.} The agent must operate within a budget, account for costs per action, and settle payments on-chain.
\end{enumerate}

No existing framework satisfies all five requirements simultaneously. AutoGPT~\cite{autogpt2023} provides persistence through file storage but no identity or economic model. LangChain~\cite{langchain2023} offers tool orchestration but treats agents as stateless pipelines. The Eliza framework~\cite{ai16z2024eliza} adds on-chain token interaction but lacks private memory and capability scoping. Virtuals Protocol~\cite{virtuals2024} tokenizes agent personas but does not address the trust kernel beneath them.

The core insight of this paper is that \emph{agent sovereignty requires the same primitives as user sovereignty}: identity, storage, capabilities, economics, and privacy. The blockchain already provides these primitives for human users. We extend them to AI agents.

We introduce the \textbf{Agent Capsule}: a self-contained sovereignty boundary that encapsulates an agent's identity, vault, capabilities, budget, and memory into a single composable unit anchored on Lux Network. The Agent Capsule is not an application-layer abstraction---it is a protocol-level primitive integrated with the Lux trust kernel.

The remainder of this paper is organized as follows. Section~\ref{sec:capsule} defines the Agent Capsule architecture. Section~\ref{sec:identity} specifies the DID-based identity model. Section~\ref{sec:memory} describes the private memory system. Section~\ref{sec:capabilities} formalizes the capability model. Section~\ref{sec:economics} presents the economic framework. Section~\ref{sec:private-inference} details private inference. Section~\ref{sec:coordination} covers multi-agent coordination. Section~\ref{sec:attestation} defines agent attestation. Section~\ref{sec:security} analyzes the security model. Section~\ref{sec:implementation} describes the implementation. Section~\ref{sec:related} surveys related work. Section~\ref{sec:conclusion} concludes.