Neither Layer Alone: Epistemic Integrity Requires Hierarchical Joint Design for Long-Running AI Agents

2026-06-04 · Source: cs.SE updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Expert, extended

Summary

Long-running AI agents fail not only when inference fails or tools are underspecified, but when independently evolving model and harness layers change the semantics of belief, capability, and goal commitments across their boundary — a failure class this paper terms "interface volatility". This paper argues that "agent epistemic integrity" (AEI) must be treated as a first-class architectural constraint, achievable only through joint model–harness design organized around an explicit interface contract. The central claim is that this contract is the precondition for joint design, operationalized by a four-level hierarchy: goal validity, action-archetype sequencing, tool-instance selection, and invocation-level failure discrimination. Evaluation and training should derive from this contract, testing whether belief, tool, and goal commitments hold across session boundaries and independent layer upgrades, rather than just local task completion.

Key takeaway

For AI Architects and Machine Learning Engineers designing long-running AI agents, you must prioritize explicit model-harness interface contracts. This prevents "interface volatility" and ensures epistemic integrity across model upgrades and session boundaries. Implement a four-level operational hierarchy for goal validity, action sequencing, tool selection, and failure discrimination. Your evaluation and training should derive from these contract obligations, not just task completion, to ensure robust, resumable agent behavior.

Key insights

Long-running AI agents need "epistemic integrity" via a hierarchical model-harness interface contract to prevent semantic drift.

Principles

• Interface volatility degrades long-running AI agents.
• Agent Epistemic Integrity (AEI) is a first-class architectural constraint.
• Explicit model-harness contracts are preconditions for joint design.

Method

A four-level hierarchy (goal validity, action-archetype sequencing, tool-instance selection, invocation-level failure discrimination) operationalizes the model-harness interface contract, replacing flat action loops with contract-preserving control over persistent state.

In practice

• Define explicit model-harness interface contracts.
• Implement a four-level operational hierarchy.
• Evaluate agents on cross-session epistemic state.

Topics

• AI Agents
• Epistemic Integrity
• Model-Harness Interface
• Interface Volatility
• Hierarchical Design
• Long-Running Systems

Best for: Research Scientist, AI Engineer, AI Scientist, AI Architect, Machine Learning Engineer

Related on AIssential

• Beyond Vibe Coding: The Artifacts Layer — Durable artifacts are essential for preserving engineering intent in agentic AI development, enabling true delegation.
• The Agent Stack - Part 2: Foundation Infrastructure, Models, and Inference — Agent system robustness depends on understanding and explicitly managing underlying infrastructure and model engine semantics.
• LAI #127: The Infrastructure Layer of AI Is Becoming the Product — Robust AI systems require careful architectural design, especially for agentic applications, to prevent silent failures and ensure reliability.
• Project NIKA: Unlocking Epistemic Agency in 4-Bit Quantized Models — Constraining LLMs with a deterministic governor can induce "epistemic agency" and superior reasoning, even in small, quantized models.
• Four-Axis Decision Alignment for Long-Horizon Enterprise AI Agents — Evaluating enterprise AI agents on four distinct alignment axes reveals nuanced failure modes hidden by aggregate metrics.
• Trust Me. I’m Artificial Intelligence. — AI models are becoming more convincing in their reasoning, but this does not equate to increased truthfulness or correctness.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.SE updates on arXiv.org.