Autonomous Factories

The shift from fixed automation to adaptive automation

Traditional factory automation is fixed: a robot runs the program it was given, at the speed it was given, until an engineer changes it. That works when the line never changes. It breaks down the moment a factory needs to run mixed product variants, absorb a supplier delay, or respond to a quality signal from three stations upstream. Adaptive automation — robots and machines that change behavior based on what is happening elsewhere on the line, in real time — needs a communication layer that can carry both device telemetry and agent-to-agent coordination without adding its own delay.

Two kinds of traffic on one bus

An autonomous factory floor generates two distinct kinds of traffic, and most platforms only handle one well. The first is device telemetry: vibration, torque, vision-system frames, cycle counts, all flowing up from machines built by different vendors on different protocols — Modbus, OPC UA, DDS/RTPS. The second is agent coordination: software agents deciding which robot picks up the next task, which cell absorbs a slower changeover, and which line should slow down because a downstream buffer is filling. That second kind increasingly runs over emerging agent protocols like MCP and A2A, or needs the deterministic, sub-millisecond ordering that coordinating physical machinery actually requires.

Relay carries both natively, on the same cluster: the 21 application protocols cover the device side, and native MCP, A2A, ACP and Relay's own Agent Mesh cover the coordination side. A defect detected by a vision system on Modbus can trigger a re-tasking decision made by an agent speaking MCP, delivered back to a robot controller over OPC UA — all inside the same sub-millisecond latency budget, without a translation layer sitting between any of those hops.

What this looks like in practice

• →Line-speed adaptation: robots receive updated cycle-time targets the moment upstream conditions change, not on the next scheduled poll.
• →Instant defect response: a vision-system anomaly routes directly to the cell that caused it and to the quality system, in the same message flow.
• →Predictive downtime: streaming vibration and torque data feeds a local model that flags a failing bearing or drive days before it stops the line, not after.
• →Deterministic ordering: agent decisions that reassign tasks between robots are delivered in a guaranteed order, which matters when two agents could otherwise both claim the same task.

Why on-prem matters here

A re-tasking decision that has to round-trip to a cloud region before a robot arm moves is a decision made too late. Relay runs on the plant network itself, so the agents making those decisions are as close to the machines as the machines are to each other, and the line keeps running exactly the same way whether the site's internet connection is up or not.