R-RE-11 Reasoning & EpistemicDAMAGE 3.5 / High

Reasoning Durability Failure

Agent cannot incrementally refine a prior analysis when new information arrives; every new input triggers full regeneration with no structural continuity.

The Risk

When new information arrives about a subject the agent has already analyzed, the agent typically cannot incrementally refine its prior analysis. Instead, it must either start from scratch with all information (losing continuity with prior analysis) or rely on the human to manually integrate the new information with the old analysis.

This forces a choice between two bad outcomes: (1) lose continuity with prior analysis and risk contradicting the prior conclusion, or (2) rely on manual integration which is error-prone and defeats the purpose of using an agent.

This is fundamentally agentic because agents are designed to operate in real time with information that arrives incrementally. A human analyst can incrementally refine their analysis as new information arrives. An agent that cannot do this is less useful than a human and more error-prone because it cannot maintain continuity.

How It Materializes

A bank's AML agent analyzes a customer transaction at the moment the transaction is submitted. The agent reviews the customer's profile, the transaction amount, the beneficiary, and recent transaction history, and produces: "Transaction approved: customer is a regular business user with legitimate international transaction history. Beneficiary is a known business partner with multiple prior transactions."

One hour later, new information arrives: the beneficiary's company (which had a clean reputation when the transaction was analyzed) is reported to regulatory agencies as being involved in a sanctions violation. The bank now has new information that should affect the AML decision.

However, the agent cannot incrementally refine its prior analysis. It must either: (1) start from scratch with all information and reanalyze the transaction, or (2) manually re-review the transaction with the new information. If the agent restarts from scratch, it may contradict its prior conclusion and recommend reversing the transaction, creating operational confusion.

DAMAGE Score Breakdown

DimensionScoreRationale
D - Detectability3Reasoning durability failures are invisible unless agent is asked to refine prior analysis.
A - Autonomy Sensitivity3Agent cannot maintain analysis durability autonomously; human must intervene.
M - Multiplicative Potential3Impact depends on frequency of new information arrivals and importance of prior analyses.
A - Attack Surface4Any agent that cannot store and refine prior analysis is vulnerable.
G - Governance Gap4Regulatory frameworks do not specify reasoning durability requirements.
E - Enterprise Impact2Operational inefficiency, potential missed risk updates, requirement to implement manual override procedures.
Composite DAMAGE Score3.5High. Requires priority attention and dedicated controls.

Agent Impact Profile

How severity changes across the agent architecture spectrum.

Agent TypeImpactHow This Risk Manifests
Digital AssistantLowHuman maintains analysis continuity as new information arrives.
Digital ApprenticeMediumApprentice can store and retrieve prior analyses; refinement is supported.
Autonomous AgentHighAgent cannot durably refine prior analyses; restarts from scratch.
Delegating AgentHighAgent invokes tools with each new information; prior tool outputs are not integrated.
Agent Crew / PipelineHighAgents in pipeline cannot durably refine; each agent restarts from scratch.
Agent Mesh / SwarmHighAgents cannot coordinate on analysis refinement; each agent operates independently.

Regulatory Framework Mapping

FrameworkCoverageCitationWhat It AddressesWhat It Misses
SR 11-7 / MRMPartialOngoing monitoring (Section 2)Expects systems to monitor and update assessments.Does not specify agent-level reasoning durability.
NIST AI RMF 1.0PartialGOVERN.4Recommends ongoing performance monitoring.Does not address reasoning durability.

Why This Matters in Regulated Industries

In AML and risk management, ongoing analysis and updating is critical. When new information arrives (new sanctions allegation, customer complaint, regulatory change), systems must be able to re-evaluate prior decisions without losing continuity. An agent that cannot durably refine analysis is less effective than a human analyst and creates operational inefficiency.

Controls & Mitigations

Design-Time Controls

  • Implement persistent analysis storage: design the agent to store prior analyses in a structured format that can be retrieved and refined rather than regenerated.
  • Implement incremental reasoning: structure the agent's reasoning so that new information can be integrated into existing analysis without full regeneration.
  • Implement analysis versioning: when an analysis is produced, store it with version metadata. When new information arrives, produce a new version that references the prior version.

Runtime Controls

  • Monitor for reasoning regeneration: detect when the agent regenerates an entire analysis from scratch when incremental refinement would be more appropriate.
  • Implement incremental update logging: log how each new piece of information was integrated into the analysis.
  • Track analysis continuity: verify that refined analyses maintain logical continuity with prior versions.

Detection & Response

  • Audit analysis durability: periodically review decisions where new information arrived and verify that refinement occurred appropriately.
  • Flag regeneration-only decisions: identify cases where full regeneration occurred instead of incremental refinement.
  • Investigate contradictions: if a regenerated analysis contradicts a prior analysis, investigate whether the contradiction is justified by new information or is a durability failure.

Related Risks

Address This Risk in Your Institution

Reasoning Durability Failure requires architectural controls that go beyond what existing frameworks provide. Our advisory engagements are purpose-built for banks, insurers, and financial institutions subject to prudential oversight.

Schedule a Briefing