Belief State: Interpreting Temporal Belief Dynamics in Agentic Financial Systems

Agentic AI systems in finance operate under uncertainty and non-stationarity, combining historical time-series signals with forward-looking disclosures to guide decisions. Existing financial agents, however, provide limited visibility or control over their internal representations of future market conditions, which constrains robustness and accountability. We present BeliefState, a mechanistic study of temporal belief representations in agentic financial models. Using earnings call transcripts aligned with historical and realized market time series, we analyze how agents construct latent belief-states that encode expectations across multiple future horizons. Through contrastive temporal probing, we identify horizon-specific internal components that respond systematically to changes in volatility regimes. We further show that these components are operationally causal through targeted activation-level interventions. Intervening on belief-states alters agent risk posture and temporal coherence during downstream reasoning. Our findings indicate that failures of financial agents under distributional shift arise in part from instability in internal belief-states rather than surface-level prediction errors. By treating temporal belief representations as inspectable system states, this work provides a practical foundation for building more robust, auditable, and responsible agentic AI systems in finance.