The Agency Collapse Hypothesis: Decoupling Narrative from Execution in LLM Agents

Autonomous LLM agents routinely fail in two seemingly distinct ways: (i) persistence failures in which the agent repeats unproductive actions under unchanged negative feedback, and (ii) state-claim failures in which the agent reports completion, verification, or success despite no corresponding verified state change. This paper proposes a unifying diagnosis: Agency Collapse, the evaluation and deployment habit of treating thought, action, and observation as a single distributional regime governed by narrative plausibility rather than by state reality. We formalize Agency Collapse as a specific instance of Subfield Collapse: distinct regimes with materially different admissibility, determinism, harm weights, and justification requirements are pooled under one interface and judged by aggregate success. Under collapse, selection pressure rewards trajectories that “close the story” cheaply, even when execution is unverified or repeated. We introduce a Split-Field Agency Stack—Thinking Field, Runtime Field, and Observation Field—coordinated by an Action Governor as an overfield that enforces invariants (read-only observation, budgeted retries, fail-closed escalation, receipt-bound justification). We specify Subfield-Stack Evaluation for Agents (SSE-A), a deployment-facing protocol that localizes failures as governable subfields and promotes mitigations only when bounded behavior is demonstrated across declared conditions. Agent reliability is reframed as stack-aware closure: agents are accepted not because their mean task score is high, but because their behavior remains bounded across the declared agency subfield stack and interface conditions.