papers/hanzo-self-improving-agents/sections/08-loop-4-maintenance-pass.tex

\section{Loop 4: Maintenance Pass}

\label{sec:loop4}

\subsection{Design Principles}

Loop~4 is deliberately human-gated. This is not a limitation but a design
choice based on three considerations:

\begin{enumerate}[leftmargin=1.5em]
    \item \textbf{Risk asymmetry.} Changes to the base tool library or system
          prompt affect all future sessions. A false positive (deploying a
          change that degrades performance) is harder to recover from than
          a false negative (delaying a beneficial change by one cycle).

    \item \textbf{Evidence requirement.} Loop~4 requires evidence from $N \geq 5$
          sessions before generating a proposal. This ensures that proposals
          are based on systematic patterns, not one-off observations.

    \item \textbf{Human expertise.} Changes to the tool library may require
          domain expertise that the LLM does not have (e.g., choosing between
          two correct implementations based on performance characteristics
          that cannot be measured in unit tests).
\end{enumerate}

\subsection{Cross-Session Aggregation}

Loop~4 aggregates telemetry across all sessions to identify systematic patterns.
The aggregation operates on the telemetry digest in each SessionReflection record,
not the raw telemetry log, to bound processing time.

\begin{definition}[MaintenanceProposal]
\label{def:maintenance_proposal}
A \textsc{MaintenanceProposal} record $\pi$ is a tuple:
\begin{equation}
\pi = \bigl( \text{id},\; \text{type},\; \text{description},\; \text{evidence},\;
             \text{expected\_impact},\; \text{risk},\; \text{status} \bigr),
\end{equation}
where:
\begin{itemize}[leftmargin=1.5em, nosep]
    \item $\text{type} \in \{\text{new-tool}, \text{modify-tool}, \text{retire-tool}, \text{system-prompt}, \text{fact-update}\}$
    \item $\text{evidence}$: References to SessionReflection records and telemetry patterns supporting the proposal.
    \item $\text{expected\_impact}$: Estimated change in tool success rate (with confidence interval).
    \item $\text{risk} \in \{\text{low}, \text{medium}, \text{high}\}$: Human-assessed risk category.
    \item $\text{status} \in \{\text{pending}, \text{approved}, \text{rejected}, \text{implemented}\}$
\end{itemize}
\end{definition}

\subsection{Proposal Generation Algorithm}

\begin{algorithm}[H]
\caption{MAINTENANCE-PASS$(\mathcal{R}_{1:N}, \mathcal{T}_{1:N})$}
\label{alg:maintenance}
\begin{algorithmic}[1]
\Require Session reflections $\mathcal{R}_{1:N}$, telemetry logs $\mathcal{T}_{1:N}$
\Ensure Set of \textsc{MaintenanceProposal} records
\State $\text{proposals} \gets \emptyset$
\State $\text{patterns} \gets \textsc{AggregateFailurePatterns}(\mathcal{T}_{1:N})$
\For{each pattern $p = (t, f, \text{count}, \text{sessions}) \in \text{patterns}$}
    \If{$p.\text{count} \geq \gamma_{\text{count}}$ \textbf{and} $|\text{sessions}| \geq \gamma_{\text{sessions}}$}
        \State $\text{supporting} \gets \{r \in \mathcal{R}_{1:N} \mid p \in r.\text{failures}\}$
        \State $\pi \gets \textsc{LLM-Propose}(p, \text{supporting}, \text{current-toolkit})$
        \State $\pi.\text{evidence} \gets \text{supporting}$
        \State $\pi.\text{status} \gets \texttt{"pending"}$
        \State $\text{proposals} \mathrel{\cup}= \{\pi\}$
    \EndIf
\EndFor
\For{each proposal $\pi \in \text{proposals}$}
    \State \textsc{HumanReview}$(\pi)$ \Comment{Blocking: awaits human approval/rejection}
    \If{$\pi.\text{status} = \texttt{"approved"}$}
        \State \textsc{ImplementProposal}$(\pi)$
    \EndIf
\EndFor
\State \Return $\text{proposals}$
\end{algorithmic}
\end{algorithm}

The thresholds $\gamma_{\text{count}} = 10$ and $\gamma_{\text{sessions}} = 3$
are defaults that balance sensitivity (detecting real problems) with specificity
(avoiding spurious proposals from noisy sessions). These thresholds should be
tuned to the agent's deployment context.

% ============================================================================
% 9. FORMAL ANALYSIS
% ============================================================================