HearthNet: Edge Multi-Agent Orchestration for Smart Homes Conference: Proceedings of the 1st ACM Conference on Agentic and AI Systems; May 26, 2026; San Jose, CA, USA Proceedings of the 1st ACM Conference on Agentic and AI Systems (CAIS ’26), May 26, 2026, San Jose, CA, USA DOI: 10.1145/3786335.3813188 ISBN: 979-8-4007-2415-2/26/05 CCS: Computing methodologies Multi-agent systems

Zhonghao Zhan, Krinos Li, Yefan Zhang, Hamed Haddadi · 2026

Read HearthNet: Edge Multi-Agent Orchestration for Smart Homes Conference: Proceedings of the 1st ACM Conference on Agentic and AI Systems; May 26, 2026; San Jose, CA, USA Proceedings of the 1st ACM Conference on Agentic and AI Systems (CAIS ’26), May 26, 2026, San Jose, CA, USA DOI: 10.1145/3786335.3813188 ISBN: 979-8-4007-2415-2/26/05 CCS: Computing methodologies Multi-agent systems by Zhonghao Zhan, Krinos Li, Yefan Zhang, Hamed Haddadi (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2604-09618.txt.

Introduction

Zhan and three colleagues build HearthNet, an edge system that lets smart-home users speak plain commands instead of hand-assembling rules, dashboards, and integrations. Real deployments break often, with failing devices, while existing agent toolkits assume short sessions. HearthNet runs persistent, role-specialized agents at the home hub, coordinating through MQTT, Git-backed shared state, and root-issued actuation leases. Planning, verification, authorization, and actuation stay separated. Three live demos cover ambiguous commands, conflict tracing, and stale-order rejection.

Catalog record

CREATORZhonghao Zhan, Krinos Li, Yefan Zhang, Hamed Haddadi
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE22008
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2604.09618

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

HearthNet: Edge Multi-Agent Orchestration for Smart Homes Conference: Proceedings of the 1st ACM Conference on Agentic and AI Systems; May 26, 2026; San Jose, CA, USA Proceedings of the 1st ACM Conference on Agentic and AI Systems (CAIS ’26), May 26, 2026, San Jose, CA, USA DOI: 10.1145/3786335.3813188 ISBN: 979-8-4007-2415-2/26/05 CCS: Computing methodologies Multi-agent systems Zhonghao Zhan, Krinos Li, Yefan Zhang, Hamed Haddadi Abstract . Smart-home users increasingly want to control their homes in natural language rather than assemble rules, dashboards, and API integrations by hand. At the same time, real deployments are brittle: devices fail, integrations break, and recoveries often require manual intervention. Existing agent toolkits are effective for session-scoped delegation, but smart-home control operates under a different scenario: it is persistent, event-driven, failure-prone, and tied to physical devices with no shared context window. We present HearthNet , an edge multi-agent orchestration system for smart homes 1 1 1 Demo video: https://www.youtube.com/watch?v=p3ZKDsKifRk ; Interactive Demo: https://hearthnet.vercel.app/ . HearthNet deploys a small set of persistent, role-specialized LLM agents at the home hub, where they coordinate through MQTT, Git-backed shared state, and root-issued actuation leases to govern heterogeneous devices through thin adapters. This design externalizes context, preserves execution history, and separates planning, verification, authorization, and actuation across explicit boundaries. Our current prototype runs on commodity edge hardware and Android devices; it keeps orchestration, state management, and device control on-premise while using hosted LLM APIs for inference. We demonstrate the system through three live scenarios: intent-driven multi-agent coordination from ambiguous natural language, conflict resolution with timeline-based tracing, and rejection of stale or unauthorized commands before device actuation. † † cc-license: by 1. Introduction AI agents are moving beyond cloud sandboxes into persistent physical environments such as home servers, IoT hubs, and personal devices. This shift exposes a mismatch between current orchestration toolkits and the realities of edge deployment. Dominant multi-agent frameworks, including Anthropic’s agent teams, Microsoft’s AutoGen, and LangChain’s LangGraph, expose state-serialization APIs but leave durable persistence and cross-machine coordination to the application. Anthropic notes that subagents remain tied to a session context window and that cross-session coordination requires a different pattern ( 3 ) . Microsoft notes that AutoGen leaves persistence and recovery to the application ( 15 ) . LangChain similarly warns that once agents become distinct graph nodes, context transfer becomes an explicit engineering problem ( 12 ) . Smart-home control operates under a different regime. Coordination is long-running, and physically consequential: devices crash, networks partition, batteries die, and stale or conflicting commands can create safety, privacy, or usability failures. These properties demand explicit shared state, recovery semantics, and permission boundaries that session-scoped subagents do not naturally provide. When orchestration spans heterogeneous devices, multiple privilege domains, and recovery from failures, a small set of long-lived specialist agents with explicit shared state is easier to audit, recover, and resume than a monolithic agent or transient subagents. We present HearthNet , an edge multi-agent orchestration system for smart homes. HearthNet deploys a small set of persistent, role-specialized LLM agents on commodity edge hardware, coordinated through MQTT, Git-backed shared state, and root-mediated authorization. The prototype keeps orchestration, state management, and device control on-premise while using hosted LLM APIs for inference. We make four contributions: • A Git-backed freshness mechanism (base commit verification) that enables agents to detect and reject stale or conflicting commands before device actuation. • Policy-bound actuation leases in which the root agent issues short-lived, machine-checkable authorizations binding each state-changing command to an identity, device scope, parameter envelope, and the state version under which it was approved. • A librarian-as-observer pattern in which a dedicated non-participant agent mirrors coordination traffic into a versioned repository, providing crash recovery, attribution, and full auditability without participating in the control path. • A deployed edge prototype running on three classes of commodity hardware (ARM64 Mac mini, x64 NUCs, Android phone), demonstrated through three live scenarios with real devices. 2. System Architecture HearthNet uses a hierarchical role architecture with four tiers (Figure 1 ): a root agent , manager agents , device adapters , and a cross-cutting librarian agent . All inter-agent coordination flows over a shared MQTT broker, while durable shared state is maintained in a Git repository managed by the librarian. A typical request begins at the root agent, which receives a user command, decomposes it into domain-specific tasks, and dispatches them to one or more managers over MQTT. Managers resolve the request against the current Git-backed state, prepare concrete device actions, and return execution requests or conflict reports. Before any state-changing operation reaches a device adapter, the root agent verifies freshness, checks policy, and issues a short-lived actuation lease. The librarian mirrors the resulting coordination traffic and records externally relevant events into the shared repository for audit and recovery 2 2 2 All code, configuration, agent templates, and replay scripts are available at https://github.com/zhonghaozhan/hearthnet . . 2.1. Roles and Communication Each major role runs as an independent edge-resident runtime, on its own device, hosted using OpenClaw ( 17 ) as the agent runtime platform. Agents coordinate over a central Mosquitto MQTT broker using point-to-point inboxes, a broadcast channel, and a mirrored audit stream. Messages are JSON-encoded and include an optional base_commit hash referencing the Git state the sender believes to be current. State-changing execution requests additionally carry a lease id referencing a root-issued authorization record (Figure 1 ). Figure 1 . HearthNet architecture. Solid arrows indicate the command path; the librarian mirrors coordination traffic and records externally relevant events to Git. Four-tier diagram showing a root agent (Rupert) on a Mac mini at the top, dispatching tasks over MQTT to two manager agents (Jeeves on an Intel NUC, Darcy on an Android phone) which control device adapters; a separate librarian agent (Dewey) on another NUC mirrors all MQTT traffic into a Git repository for audit and recovery. Root Agent. "Rupert" on a Mac Mini serves as the system’s top-level orchestrator. It receives high-level user commands via Telegram, decomposes them into domain-specific tasks, dispatches them to managers, arbitrates conflicts, and is the only component allowed to issue actuation leases. Rupert holds the broadest system context but delegates device-specific resolution to managers. Manager Agents. Managers each govern a domain of devices and report to Rupert: • Home-Assistant / NUC Manager ( "Jeeves" ). Runs on an x64 Intel NUC and manages devices exposed through Home Assistant ( 10 ) , including lights, switches, climate, and media endpoints. • Mobile-App / Phone Manager ( "Darcy" ). Runs on a Pixel 4 (Android 14) phone and controls proprietary IoT applications that expose no external API, using UI automation through ADB and accessibility services. Darcy also accesses on-device sensors (camera, ambient light, gyroscope) that are physically unavailable to hub-resident agents, a key reason a single centralized agent cannot replace the distributed architecture. Managers can propose device actions within their domain, but they do not hold standing authority to actuate devices; every state-changing command requires a fresh lease from Rupert. Librarian Agent. "Dewey" runs on a separate Intel NUC and maintains a Git repository as the canonical shared record of system coordination state. It subscribes to mirrored MQTT traffic and records externally relevant events with structured metadata (timestamp, sender, outcome). The events include task dispatches, device-state transitions, conflict

Read the complete text