mirror of
https://github.com/hanzoai/papers.git
synced 2026-11-03 02:27:30 +00:00
53 lines
2.4 KiB
TeX
53 lines
2.4 KiB
TeX
\section{Related Work}
|
|
|
|
\label{sec:related}
|
|
|
|
\subsection{Agent Frameworks}
|
|
|
|
Early agent frameworks including LangChain~\cite{langchain2022} and
|
|
AutoGPT~\cite{autogpt2023} established the pattern of wrapping LLM inference
|
|
with tool-use capabilities. These frameworks provide useful abstractions but do
|
|
not address compute provisioning, multi-channel identity, or observability as
|
|
first-class concerns. The OpenAI Assistants API~\cite{openai2023gpt4} provides
|
|
managed tool execution but ties the agent runtime to a single model provider,
|
|
creating the vendor lock-in problem described in Section~\ref{sec:intro}.
|
|
|
|
\subsection{Coding Agents}
|
|
|
|
SWE-bench~\cite{jimenez2024swebench} established a rigorous benchmark for
|
|
software engineering agents, catalyzing a wave of systems including
|
|
Devin~\cite{cognition2024devin}, Aider~\cite{aider2024}, and SWE-
|
|
agent~\cite{yang2024sweagent}. These systems demonstrate that LLMs can resolve
|
|
real GitHub issues when equipped with appropriate tools and scaffolding.
|
|
However, they treat the execution environment as a fixed background assumption
|
|
rather than a dynamically managed resource.
|
|
|
|
\subsection{Computer Use}
|
|
|
|
Anthropic's computer use capabilities~\cite{anthropic2024computer} and
|
|
OpenAI's Operator~\cite{openai2025operator} enable LLMs to interact with
|
|
graphical interfaces. The Hanzo Operative framework~\cite{hanzo2024operative}
|
|
extends this with formal action specifications and safety boundaries. These
|
|
systems occupy a single layer of the UAH architecture---the compute tier---
|
|
without integrating the surrounding infrastructure.
|
|
|
|
\subsection{Model Context Protocol}
|
|
|
|
The Model Context Protocol (MCP)~\cite{mcp2024} defines a standard interface
|
|
for tool discovery and invocation, analogous to USB for AI tools. MCP
|
|
standardizes the tool protocol layer but leaves all other layers unspecified.
|
|
The UAH adopts MCP as the tool protocol standard while adding six additional
|
|
layers above and below.
|
|
|
|
\subsection{Multi-Agent Systems}
|
|
|
|
Research on multi-agent coordination~\cite{park2023generative,
|
|
wu2023autogen} has demonstrated emergent capabilities from collections of
|
|
specialized agents. The UAH provides the routing, shared state, and
|
|
observability infrastructure that makes multi-agent coordination tractable at
|
|
production scale.
|
|
|
|
% ============================================================================
|
|
% 3. THE SEVEN-LAYER ARCHITECTURE
|
|
% ============================================================================
|