mirror of
https://github.com/hanzoai/papers.git
synced 2026-10-29 07:34:31 +00:00
62 lines
3.3 KiB
TeX
62 lines
3.3 KiB
TeX
\section{Related Work}
|
|
|
|
\label{sec:related}
|
|
|
|
\paragraph{Tool-Creating Agents.}
|
|
The closest prior work is \textsc{bmo}~\cite{ngrok2024bmo}, an agent that
|
|
can generate and register new tools at runtime to address capability gaps.
|
|
\textsc{bmo} demonstrates that tool creation is practically viable but does not
|
|
address extension development, cross-component modification, or the formal safety
|
|
properties required for production deployment. Our work extends \textsc{bmo}'s
|
|
runtime creation to a full four-tier hierarchy with isolation guarantees.
|
|
|
|
\paragraph{Skill Libraries.}
|
|
Voyager~\cite{wang2023voyager} builds an ever-growing skill library for
|
|
Minecraft by having GPT-4 write JavaScript code that is verified and stored for
|
|
reuse. Each skill is tested in-game before storage. HHP generalizes this idea
|
|
beyond game-specific code to production software components, and adds the
|
|
critical missing piece: a safety framework for when the skill is not a
|
|
sandboxed game action but a modification to shared production infrastructure.
|
|
|
|
\paragraph{Automated Design of Agentic Systems.}
|
|
ADAS~\cite{hu2024adas} uses a meta-agent to search the space of agentic system
|
|
designs, producing new agent specifications expressed in code. ADAS operates
|
|
at design time, not at agent runtime. HHP addresses the complementary problem:
|
|
runtime self-modification during active task execution, where modifications must
|
|
not disrupt the running session.
|
|
|
|
\paragraph{Self-Instruct and Self-Improvement.}
|
|
Self-Instruct~\cite{wang2022self} and related work~\cite{zelikman2022star,
|
|
yuan2023scaling} demonstrate that models can generate training data and
|
|
instructions that improve themselves. These methods improve model weights;
|
|
HHP improves the harness rather than the weights, making it complementary to
|
|
and composable with weight-based self-improvement.
|
|
|
|
\paragraph{Program Synthesis and Self-Modifying Code.}
|
|
Program synthesis~\cite{gulwani2017program} generates code from
|
|
specifications. Classical self-modifying code~\cite{holland1975adaptation}
|
|
allows programs to rewrite their own instructions. Modern Genetic
|
|
Programming~\cite{koza1994genetic} evolves programs via mutation and selection.
|
|
HHP draws on these traditions but differs in that modifications are semantically
|
|
motivated (derived from observed friction), architecturally structured (four
|
|
tiers), and formally safety-bounded rather than evolutionary.
|
|
|
|
\paragraph{Git-Based Review as Safety.}
|
|
The use of pull request review as a human gate for agent-generated code
|
|
modifications has been explored in the context of automated bug
|
|
fixing~\cite{sobania2023analysis} and automated refactoring~\cite{bavishi2019autopandas}.
|
|
HHP systematizes this pattern: the PR is not merely a record but a required
|
|
gate for all Tier 4 (cross-component) modifications, with CI as an automated
|
|
co-reviewer.
|
|
|
|
\paragraph{Active Semantic Optimization.}
|
|
ASO~\cite{hanzo2026aso} and DSO~\cite{hanzo2026dso} provide the telemetry and
|
|
experience-sharing infrastructure that HHP builds on. GRPO-extracted experiential
|
|
priors in ASO encode what approaches work best for a given agent; HHP closes the
|
|
loop by allowing those insights to drive harness modifications that make the
|
|
winning approaches available as first-class tools.
|
|
|
|
% ============================================================================
|
|
% 3. THE SELF-MODIFICATION TAXONOMY
|
|
% ============================================================================
|