A single large language model call cannot reconcile the fragmented knowledge that a real enterprise holds across databases, documents, and line-of-business systems, and a standard retrieval-augmented generation (RAG) pipeline over that same knowledge base still returns one point of view per query. Independent multi-agent designs help by giving each business domain its own agent, but without a coordination layer the agents duplicate effort and disagree, which quietly lowers the reliability of the decision that reaches a human. This paper presents AgentMesh-AI, a seven-layer framework built around an Adaptive Agent Coordination Engine (AACE) that routes enterprise tasks to domain-specialized agents, and a Knowledge Flow Reliability Score (KFRS) that scores how trustworthy the resulting decision is before it is shown to a user. The framework is evaluated on a seeded, paired-task benchmark harness spanning four enterprise scenarios — business-intelligence analytics, financial risk, supply-chain disruption, and compliance review — against a single-LLM baseline, a standard RAG baseline, and an uncoordinated multi-agent baseline (160 paired tasks per system). AgentMesh-AI raises pooled decision accuracy from 60.8–68.3 % to 78.7 % and KFRS from 55.9–63.8 to 80.2, with Wilcoxon signed-rank tests confirming the gap over every baseline (p < .001, rank-biserial r = 0.66–0.87). Because the evaluation uses a parameterized simulation rather than live production telemetry, the numbers are best read as evidence of relative ordering under controlled conditions rather than an absolute performance guarantee, and we discuss what a production pilot would still need to confirm.

Multi-Agent Systems,Enterprise Decision Intelligence,Knowledge Orchestration,Retrieval-Augmented Generation,Agent Coordination,Knowledge Flow Reliability Score,Large Language Models,Intelligent Computing