A Multi-agent Orchestration Architecture for Reliable Task Execution in Large Language Model Systems
Year: 2026
Type: article
Abstract: The rapid deployment of large language models (LLMs) into production environments has exposed a fundamental reliability gap: a single monolithic model prompted to plan, execute, and verify a complex task in one pass frequently produces plausible but incorrect results, with no internal mechanism to detect its own failures. This study proposes a multi-agent orchestration architecture in which a complex task is decomposed across specialized LLM agents — a Planner, one or more Executors, and a Critic — coordinated by a deterministic controller that manages state, routing, and termination. Each agent operates with a constrained role, a temperature of zero, and a structured output schema validated against a typed contract, allowing the controller to reason about agent outputs programmatically rather than through free-form text parsing. The proposed controller treats task execution as a finite-state process: the Planner produces an explicit task graph, Executors resolve individual sub-tasks, and the Critic evaluates each result against acceptance criteria, triggering bounded retries or re-planning when validation fails. Experimental evaluation on a benchmark of 300 synthetic multi-step tasks shows that the orchestrated multi-agent system achieves a task-success rate of 0.91, compared with 0.68 for a single-agent baseline and 0.79 for a single agent augmented with chain-of-thought prompting. The orchestrated system also reduced unrecoverable failures by 71% through critic-triggered retries. Results indicate that explicit role separation, structured inter-agent contracts, and deterministic control flow provide a practical and auditable foundation for reliable LLM-based automation.
Author: Komronbek H. Obloev
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W7220263312
