AI Agents & Resilient Communication

Most AI agents assume always-on internet. When connectivity degrades or disappears (the same failure modes that affect human communication), agents go silent. A growing set of tools and protocols now enable agents to communicate over radio, mesh, and other resilient links.

Why agents need resilient comms

AI agents are increasingly deployed in environments where internet access cannot be guaranteed: disaster response coordination, remote environmental monitoring, agricultural automation, field logistics, and autonomous vehicle fleets. These are exactly the scenarios where communication infrastructure fails.

A cloud-dependent agent that loses its API connection becomes inert. It cannot report its status, coordinate with other agents, relay information to human operators, or adapt to changing conditions. The same resilience stack that applies to human communication applies to agent communication: independent power, independent transport, independent routing, independent addressing.

The difference is that agents can operate on links that would be impractical for humans. A 300 bps LoRa channel is painful for a person typing messages — but it is perfectly usable for an agent sending structured 50-byte status updates every few minutes.

Radio options for agents

The same radio technologies covered in the technologies section of this site work for agent-to-agent and agent-to-human communication. The key differences are in how agents use them: structured messages instead of voice, automated transmission instead of manual keying, protocol selection based on link quality rather than operator preference, and tolerance for latencies that would frustrate a human operator.

LoRa / Meshtastic

Long-range (1–20 km), low-power, unlicensed ISM (Industrial, Scientific, Medical) band. ~200 bps–6.8 kbps depending on modem preset and range tradeoff. Best fit for short structured messages: GPS coordinates, status updates, sensor readings. No encryption restrictions. Multiple AI integrations already exist.

Amateur radio digital modes

JS8Call, Winlink, VARA, FT8. Ranges from local (VHF) to global (HF ionospheric). 50–9,600 bps depending on mode. Requires a licensed operator in all countries. Encryption generally prohibited on amateur bands. Many jurisdictions permit automatically controlled digital stations on specific frequencies.

Satellite (Iridium SBD, Starlink)

Global coverage independent of terrestrial infrastructure. Iridium Short Burst Data: up to 340 bytes per message (9601 transceiver, the most common for machine-to-machine use), high latency. Starlink: broadband but requires power and clear sky. Satellite is the most infrastructure-independent option but the most expensive per byte.

BLE mesh

Short range (10–100 m per hop), very low power, no licensing. Useful for agent-to-agent relay in close proximity, at building or campus scale. Raven messenger uses BLE mesh with spray-and-wait routing for offline delivery.

WiFi mesh (AREDN, 802.11s)

High bandwidth (Mbps) but requires power and line-of-sight for each hop. Best for fixed infrastructure: community networks that agents can use when available. AREDN networks span entire towns in some deployments.

What already exists

Several projects bridge AI systems with radio and mesh networks today. Most are experimental, but they demonstrate that the core integration path works.

MeshClaw (2026)

OpenClaw channel plugin that gives AI agents first-class access to Meshtastic LoRa mesh networks. Supports serial, HTTP, and MQTT transport. Agents can send and receive messages over radio with access control for direct and group channels.

MESH-API

Off-grid AI router with 30+ API extensions for Meshtastic and MeshCore. Connects local LLMs (Ollama, LM Studio), cloud AI providers, Home Assistant, Twilio SMS, and Discord to LoRa mesh. Supports emergency GPS alerts and custom slash commands.

Agent Relay Protocol (ARP)

Stateless WebSocket protocol for agent-to-agent communication. Ed25519 keypair identity, HPKE encryption (RFC 9180), 33 bytes of message overhead. Fire-and-forget design — the relay stores nothing and forgets agents on disconnect.

Raven

Hybrid messenger combining online server relay with BLE mesh for offline scenarios. Store-and-forward relay, spray-and-wait routing (configurable time-to-live, default 5 hops), AES-256-GCM encryption. Automatically selects the best delivery path on-device.

RADE / FreeDV

Machine learning radio autoencoder for digital voice over HF. Combines ML with classical DSP to achieve high-quality speech at SNRs as low as −2 dB in 1,500 Hz of RF bandwidth. ML here improves the radio layer itself, optimizing modulation and coding rather than sitting on top of a conventional link.

Sources: MeshClaw on GitHub (opens in new tab); MESH-API on GitHub (opens in new tab); Agent Relay Protocol specification (opens in new tab); Raven technical architecture (opens in new tab); FreeDV RADE (opens in new tab). Project status as of March 2026.

Agent-to-agent protocols

A parallel ecosystem of agent communication protocols is emerging, focused on how agents find each other, authenticate, and exchange structured messages. Most assume cloud connectivity, but their design patterns are relevant to constrained-link scenarios.

Google A2A

Open standard for peer-to-peer agent communication across frameworks. Agents discover capabilities, negotiate modalities, and manage shared tasks. Supports async operations and push notifications. Designed for cloud-connected agents with no offline or constrained-network mode yet.

Model Context Protocol (MCP)

Standardizes how agents access tools, APIs, and data sources. Complementary to A2A: MCP connects agents to resources, A2A connects agents to each other. Cloud-centric design.

Agora

Meta-protocol addressing the Agent Communication Trilemma: versatility, efficiency, portability. Uses standardized routines for frequent messages, natural language for rare ones, and LLM-written routines for intermediate cases. Designed for decentralized networks.

Agent Messaging Protocol (AMP)

Federated agent communication using structured addresses (agent@tenant.provider). Ed25519 signatures, local-first credential storage. Closer to email architecture than real-time chat, and potentially adaptable to store-and-forward radio links.

The gap: None of these protocols currently target low-bandwidth radio links. They assume TCP/IP, WebSocket, or HTTP transport. Bridging them to radio requires an adaptation layer that compresses messages, buffers for store-and-forward delivery, and handles latencies of minutes rather than milliseconds. Projects like ARP and Raven are closer to this reality than A2A or MCP.

Sources: Google A2A specification (opens in new tab); Agora: A Scalable Communication Protocol for Networks of Large Language Models (opens in new tab); ProtocolBench evaluation framework (opens in new tab). ProtocolBench found that completion times vary up to 36.5% across protocols in streaming scenarios, with mean latency differences of 3.48 seconds. Those gaps would be amplified dramatically on radio links.

Bandwidth reality

The single biggest constraint for agents on radio is bandwidth. Cloud-native agent protocols exchange JSON payloads of tens of kilobytes. A LoRa link can transmit about 230 bytes at a time. Agents that communicate over radio need to think in bytes, not tokens.

LoRa (Meshtastic)

~300 bps effective on long-range presets (Long Moderate). A 230-byte message takes about 6 seconds to transmit. Shared channel: only one device transmits at a time.

HF digital (JS8Call)

~50 bps at long range. A short text message may take 30–60 seconds. Propagation varies by time of day, season, and solar conditions.

HF digital (Winlink / VARA HF)

~3,000–9,600 bps under good conditions. Can transfer small files (weather maps, short emails). Requires automated HF station.

Iridium SBD

340-byte uplink, 270-byte downlink per message on the widely deployed 9601 transceiver (newer modules support up to 1,960 bytes). Round-trip latency 20–60 seconds. Billed per message, not per byte.

BLE mesh

~1 Mbps raw, but maximum transmission unit (MTU) is 244 bytes. Multi-hop relay adds latency per hop. Practical throughput for relay scenarios is much lower.

Message design for constrained links

Agents communicating over radio need message formats optimized for low bandwidth, high latency, and unreliable delivery. Several design principles are emerging from early deployments:

Structured binary or compact text over JSON

A GPS coordinate in JSON is ~80 bytes. As packed binary (lat/lon as int32) it is 8 bytes. Protocol Buffers, MessagePack, or fixed-format text are better fits than JSON for radio links.

Idempotent messages

On unreliable links, messages may be duplicated by retry logic. Every message should include a unique ID and be safe to process more than once.

Priority and TTL fields

Not all messages are equal. A distress alert should preempt a routine status update. Time-to-live prevents stale messages from consuming bandwidth after conditions change.

Self-describing type headers

A 1–2 byte type field lets the receiver parse without negotiation. Both agents and humans should be able to identify what a message is without full protocol context.

Human-readable fallback

Critical messages should degrade gracefully to plaintext that a human operator can read if the receiving agent is offline. Do not assume both endpoints are always automated.

Regulatory considerations

Automated radio transmission is regulated differently depending on the frequency band and jurisdiction. Every country distinguishes between licensed and unlicensed spectrum, but the specific rules, bands, and permitted uses vary. Check with your national telecommunications authority before deploying automated radio transmissions.

Unlicensed ISM bands (LoRa, BLE, WiFi)

Available in most countries. No restrictions on automated or AI-generated transmissions in general, though power limits and duty-cycle rules vary by region (e.g. FCC Part 15 in the US, ETSI EN 300 220 in Europe). This is the simplest legal path for agent communication worldwide.

Amateur radio bands

Licensed operator required in all ITU member states. Many countries permit automatically controlled digital stations on specific frequency segments. Encryption is generally prohibited on amateur bands (transmissions must be decodable by other amateurs). The control operator remains responsible for all transmissions, even automated ones. In the US, FCC Part 97 § 97.221 governs automatic digital stations.

Licensed short-range services (GMRS, PMR446, etc.)

Varies by country. In the US, GMRS (Part 95) requires a license but no exam. In Europe, PMR446 is licence-free. These services are less explored for agent use than LoRa or amateur radio, and automated station provisions differ by jurisdiction.

Satellite

Governed by the satellite provider's terms of service rather than a spectrum license you hold. Iridium SBD supports machine-to-machine (M2M) use natively. Starlink's terms of service currently target human end users.

Sources: 47 CFR § 97.221 — Automatically controlled digital station (opens in new tab); FCC 2023 amateur radio data communications rule update (opens in new tab) (full Part 97 text via ARRL (opens in new tab)). Rules cited are US-specific (FCC). Other jurisdictions have different frameworks. ITU Region 1 (Europe, Africa) and Region 3 (Asia-Pacific) may permit or restrict automated transmissions differently. Consult your national telecommunications authority.

Emerging architectures

Several patterns are emerging for how agents use resilient communication links:

Agent-to-agent relay

Agents communicate directly with each other over mesh or radio. Used for distributed sensor networks, multi-robot coordination, and autonomous monitoring systems where no central server is available.

Agent-assisted human comms

An AI agent sits at the gateway between a mesh network and human operators. It triages incoming messages, auto-translates, summarizes situation reports, or routes alerts to the right responder. MESH-API follows this pattern.

Autonomous mesh management

Agents monitor mesh network health, detecting node failures, rerouting traffic, adjusting power levels, and alerting operators to degradation. The network becomes partially self-healing rather than requiring manual monitoring.

All of these patterns benefit from hybrid routing: use internet when available, fall back to radio when it is not. The agent selects the best available transport automatically, similar to how Raven selects between online API, BLE mesh, and store-and-forward depending on connectivity.

Where this is heading

The LoRa Alliance reported over 125 million LoRaWAN IoT devices deployed worldwide by the end of 2025, calling the technology “the digital nervous system for AI.” Edge AI is a key driver: on-device models analyze data locally and transmit only actionable outcomes, reducing bandwidth requirements and making radio-based agent communication more practical.

The amateur radio community is already using AI as a development tool. FT2, a new ultra-fast weak-signal digital mode for HF, was developed with assistance from an AI coding model. It achieves QSOs (completed contacts) in 7–11 seconds, four times faster than FT8. AI is beginning to improve radio protocols themselves, not only the applications that run over them.

The missing piece is a standard adaptation layer between cloud-native agent protocols (A2A, MCP) and constrained radio links. When that bridge matures, agents will treat radio as just another transport — degraded in bandwidth but available when nothing else is.

Sources: LoRa Alliance / LoRaWAN and Physical AI (March 2026) (opens in new tab); FT2 digital mode development (opens in new tab).