Efficient and Reliable LLM Agents for Collaborative Intelligent Decision-Making

Abstract

Large language model agents have moved from prompt-centric assistants to tool-using, memory-augmented, multi-agent systems that must operate under real constraints: bounded inference budgets, long-horizon reasoning, privacy and policy requirements, and domain-specific decision objectives. First, evaluation still dominates the field: many papers define new benchmarks for telecommunications, asynchronous planning, clinical workflows, economic environments, API nesting, customer support, wealth management, and financial document processing. Second, reasoning quality is increasingly being treated as a systems problem rather than a prompting problem alone; memory, curriculum design, calibration, and compute scheduling are all used to trade cost, latency, and reliability against task success. Third, the move from single-agent prompting to multi-agent orchestration introduces new gains in role specialization and tool use, but also new risks in information leakage, collusion, policy evasion, and judge reliability. The review concludes that the literature is converging on a common design principle: effective agentic intelligent systems are not only about stronger models, but about co-design across inference control, reasoning structure, collaboration protocols, and auditable decision workflows. The paper also includes corpus-level experiments over the provided references, several comparative tables, and synthesis figures.

Cover Image
Downloads
Published
2026-08-01
Section
Articles