"Agentic Decentralized Autonomous Machines Dataset"

"Decentralized Physical Infrastructure Networks (DePINs) often rely on static, rule-based protocols that degrade under sensor drift, noise, and intermittent connectivity, and offer limited support for coordinated, context-aware decisions across distributed IoT devices. This paper presents \\emph{Agentic Decentralized Autonomous Machines} (ADAM), a crew-based multi-agent architecture in which edge nodes form task-specific teams on demand, share context via semantic memory, selectively invoke structured LLM inference through an API, and enforce policy validation with immutable audit logging on a Proof-of-Authority (PoA) blockchain. ADAM follows a hybrid edge--cloud design: edge devices perform continuous sensing and event-driven coordination, while semantic interpretation is offloaded only when needed. On a methane monitoring testbed, ADAM improves $F_1$ by 13.1 percentage points over a static-threshold baseline, while sustaining a mean end-to-end decision latency of 1.26\\,s over a 72-hour deployment and operating at \\emph{25.1}\\% average CPU utilization on Raspberry Pi--class devices. These results indicate that crew-based, LLM-assisted coordination with governance-backed validation can improve robustness in non-stationary edge monitoring without exceeding practical resource budgets."