Structured Elicitation Primitives for Reliable Multi-Agent Delegation and Recursive Planning

Abstract

The transition of Large Language Models (LLMs) from passive information retrieval interfaces to agentic systems capable of multi-step execution represents a significant paradigm shift in artificial intelligence. However, the reliability of these agents is frequently compromised by stochastic drift, hallucination, and the inability to maintain coherent context over extended planning horizons. This paper proposes a theoretical framework for Reliable Agent Delegation (RAD), focusing on structured elicitation techniques that constrain the probabilistic output of foundation models into deterministic workflows. We analyze role assignment mechanisms, meta-reasoning prompts, and self-corrective failure recovery loops. Drawing upon existing literature in Chain-of-Thought (CoT) reasoning, ReAct frameworks, and formal verification, we posit that imposing rigid syntactic and semantic constraints on elicitation allows for verifiable delegation between orchestrator and worker agents. We discuss the security implications of such architectures, specifically regarding indirect prompt injection and cascading logic failures, and outline a methodology for constructing robust, self-healing agentic systems.