Get the NATS Newsletter
News and content from across the community
AI is becoming a mesh of specialized agents.
The industry is moving past single general-purpose assistants toward meshes of specialized agents, each with a defined role, domain context, and operational responsibility.
The bottleneck isn't the model. It's the system around it.
As the number of agents grows, so does the need for reliable, real-time coordination between them. The winning AI systems will be defined by their connective infrastructure, not their model.
The highest-stakes AI will operate in the physical world.
When agents control energy grids, manufacturing lines, and vehicles, intelligence becomes even more impactful — but there's no rollback. Every decision must be grounded in real-world, operational context.
Events turn intelligence into action.
In live environments, agents won't wait to be prompted. They'll act on schedules, heartbeats, state changes, and environmental signals. They'll be proactive participants in the systems they operate, not passive responders behind a chat interface.
Four capabilities, each building on the last.
NATS connects AI to telemetry, command, and control loops in real time, so agents can observe and intervene, not just analyze after the fact.
# Subscribe to sensor telemetry
nats sub "factory.line3.sensors.>"
# Request current status from any agent
nats req "agent.hvac.status" ""
# Publish a control command
nats pub "factory.line3.cmd.halt" \
'{"reason":"temp_threshold"}'Event-driven systems are the natural foundation for multi-agent AI. With subject-based routing and JetStream persistence, agents activate on the signals that matter to them, act independently, and coordinate across domains.
# Create a persistent event stream
nats stream add FLEET_EVENTS \
--subjects "fleet.*.events.>"
# Route agent consumes its domain
nats consumer add FLEET_EVENTS route_agent \
--filter "fleet.*.events.deviation"
# Agent publishes decision for peers
nats pub "fleet.v1042.decision.reroute" \
'{"depot":"west","eta":"14m"}'In operational environments, answers are assembled, not retrieved. Agents pull live state, correlate streams, check external inputs, and assemble short-lived programs to decide what happens next.
# Agent queries live state from three services
nats req "grid.sub.*.load" ""
nats req "api.weather.solar" \
'{"region":"northeast"}'
nats req "ops.maintenance.check" \
'{"window":"current"}'
# Agent publishes its decision
nats pub "grid.decision.rebalance" \
'{"action":"shed_15mw","conf":0.94}'NATS brings the same primitives for eventing, coordination, resilience, and security across distributed environments. With leaf nodes, gateways, and native security, agents stay connected to hubs when they can, but remain operational when they can't.
# Edge node connects to cloud cluster
leafnode {
remotes [{
url: "tls://cloud.synadia.com"
credentials: "./edge-node.creds"
}]
}
# Local agents keep running if
# cloud connection drops
# Resync automatically on reconnectUnderstanding how real-time data infrastructure powers the next generation of AI systems — from edge to cloud.
News and content from across the community