papers/hanzo-aso/sections/08-hanzo-dev-agent-architecture.tex

\section{Hanzo Dev Agent Architecture}

\label{sec:agent}

\subsection{System Overview}

Hanzo Dev is a production CLI agent that implements ASO for agentic code
generation. It integrates with standard development workflows and provides a
simple interface for automated software engineering tasks.

\subsection{CLI Interface}

\begin{verbatim}
hanzo dev solve <issue_file>
  --repo <path>              # Target repository
  --commit <hash>            # Git commit to base from
  --group-size 4             # TF-GRPO group size (G)
  --max-iterations 3         # Maximum refinement iterations (K)
  --test-cmd "pytest"        # Test command for reward signal
  --prior-bank <path>        # Path to prior bank (LanceDB)
  --temperature 0.7          # PoE decoding temperature
  --epistemic-weight 0.3     # Weight for exploration term (beta)
\end{verbatim}

\subsection{Workflow Pipeline}

The agent executes the following pipeline for each issue:

\begin{enumerate}[leftmargin=1.5em]
    \item \textbf{Issue Analysis:} Parse issue description, extract requirements,
          identify affected components via semantic search over the codebase.
    \item \textbf{Code Localization:} Use embedding-based retrieval to identify
          relevant files and functions. Rank by relevance score and filter by
          edit distance to the issue description.
    \item \textbf{Prior Retrieval:} Query the prior bank for experiential priors
          relevant to the current issue type (content-addressed by task
          embedding similarity).
    \item \textbf{TF-GRPO Rollouts:} Generate $G$ candidate patches using PoE
          decoding with retrieved priors. Each rollout produces a complete
          git-compatible patch.
    \item \textbf{Test Execution:} Apply each patch, run the test suite, collect
          pass/fail results and coverage metrics.
    \item \textbf{Advantage Extraction:} Compute group-relative advantages,
          extract semantic patterns via LLM introspection.
    \item \textbf{Prior Compression:} Distill patterns into 1-bit expert factors,
          store in LanceDB with Merkle proofs.
    \item \textbf{Iteration or Termination:} If no solution passes all tests
          and $k < K$, return to step 4 with updated priors. Otherwise, select
          the best patch and emit it.
\end{enumerate}

\subsection{Prior Bank Storage}

Experience priors are stored in a local LanceDB instance with the following schema:
\begin{itemize}[leftmargin=1.5em]
    \item \textbf{Key:} SHA-256 hash of the task specification embedding.
    \item \textbf{Value:} 1-bit quantized delta vector + per-expert scale factor.
    \item \textbf{Metadata:} Timestamp, quality score $q_m$, source attribution,
          iteration number, task type embedding.
    \item \textbf{Integrity:} Merkle proof for optional on-chain registry
          submission (see DSO protocol).
\end{itemize}

The prior bank supports efficient nearest-neighbor queries over task embeddings,
enabling retrieval of relevant priors for new tasks in $O(\log N)$ time where
$N$ is the total number of stored priors.

% ============================================================================
% 9. EXPERIMENTAL RESULTS
% ============================================================================