papers/hanzo-federated-agents/sections/10-discussion-and-future-work.tex

\section{Discussion and Future Work}

\label{sec:discussion}

\paragraph{PoAI Integration.}
Proof-of-AI (ZIP-002)~\cite{hanzo2026poai} provides verifiable attestation
that experiences were genuinely extracted from real agent trajectories, not
fabricated. Integrating PoAI with FAI would provide a cryptographic
certification layer, allowing nodes to assign higher quality scores to
PoAI-attested experiences. This addresses the scenario where Byzantine nodes
submit syntactically valid but semantically fabricated patterns that pass
the DeltaSoup filter.

\paragraph{Hierarchical DSO for Large Fleets.}
For very large fleets ($N > 1000$), flat gossip broadcast becomes bandwidth-
intensive. A hierarchical protocol organizes nodes into clusters
(cluster $\to$ region $\to$ global), with DeltaSoup applied at each level:
\begin{equation}
    \mathcal{L}_{\text{global}} = \textsc{DeltaSoup}\!\left(
        \left\{\textsc{DeltaSoup}(\{\mathcal{L}_{i,r}\}_{i \in R}) : r \in \mathcal{R}\right\}
    \right),
\end{equation}
where $\mathcal{R}$ is the set of regions and $\{i \in R\}$ denotes nodes in
region $R$. The two-level Byzantine robustness guarantees are preserved as
long as each region has $f < N_r / 3$ Byzantine nodes.

\paragraph{Economic Incentives for Experience Sharing.}
High-quality experience contributions should be rewarded to incentivize
participation. An integration with the DSO token economy would issue \$AI
tokens to nodes whose contributed experiences (as measured by the improvement
they produce in receiving nodes' task success rates) exceed a quality
threshold. This requires a mechanism for attributing outcome improvement to
specific received experiences---a causal inference problem that remains open.

\paragraph{Cross-Domain Transfer via Embedding Alignment.}
Currently, domain filtering restricts cross-domain sharing. For tasks that
span multiple domains (e.g., a coding task that also requires documentation),
embedding alignment techniques~\cite{conneau2020unsupervised} could map
experiences from different domains into a shared semantic space, enabling
more flexible knowledge transfer.

\paragraph{Online GRPO.}
The current system runs GRPO in batch mode after sessions conclude. An
online variant would update the experience library in real time during
agent execution, allowing rapid adaptation within a single long session.
This requires careful management of the GRPO group buffer and may interact
with active DeltaSoup merges.

\paragraph{Adversarial Pattern Detection.}
DeltaSoup's geometric median filter catches embedding-space adversaries
but may not detect experiences with valid embeddings that carry subtle
misinformation in their natural language text (e.g., patterns that appear
useful but subtly suggest insecure practices). Future work should incorporate
semantic consistency checks: verifying that the pattern text is consistent
with the embedding, and that applying the pattern on held-out tasks does not
degrade performance.

% ============================================================================
% 11. RELATED WORK
% ============================================================================