Epistemic Integrity: New Framework for Reliable Long-Running AI Agents

A new research paper from Zhihong Shen, published on arXiv in June 2026, identifies a critical failure mode for long-running AI agents that goes beyond simple inference errors or underspecified tools. The paper coins the term 'Interface Volatility' to describe what happens when the model layer (which handles reasoning and beliefs) and the harness layer (which manages tools and state) evolve independently, causing semantic mismatches in belief, capability, and goal commitments. The author argues that existing flat action loops cannot guarantee reliability over extended sessions, and that this problem will only intensify as agents operate for hours, days, or weeks without human intervention.

To address this, Shen proposes that Agent Epistemic Integrity (AEI) must be treated as a first-class architectural constraint. The key insight is that the model-harness interface must be governed by an explicit contract organized into four hierarchical levels: goal validity (ensuring the agent's objectives remain coherent), action-archetype sequencing (defining permissible sequences of high-level actions), tool-instance selection (specifying which tools and instances are appropriate for each archetype), and invocation-level failure discrimination (distinguishing between recoverable and terminal errors). This hierarchy provides a structured output specification that the model must return, allowing the harness to verify integrity across sessions and independent upgrades. The paper reframes long-running agent design away from flat action loops and toward contract-preserving control over persistent state, with evaluation and training deriving from the contract itself rather than ad-hoc metrics.