How the AgentMesh Sovereign Pod reads machine signals, decides inside a 30-millisecond budget, writes the work order, and improves every night, with the safety PLC kept in charge and no data leaving the site.
AgentMesh is a small pod of servers on the factory network, next to the robots, test rigs, battery lines and tools it watches. It runs one loop in four stages. Sense reads signals without writing back to them. Decide judges each event in milliseconds. Reason writes up what happened and what to do about it. The Kiln reviews the day every night and makes small, tested improvements. AgentMesh acts on its own only where the consequence is low. Anything touching process safety goes to the PLC, the safety system and the operator as an alarm and a request. Numbers marked "target" have not yet been measured on a customer line.
New factories are full of sensors, robots and AI-driven tools, and they produce faster signals than people can watch. AgentMesh sits next to the machines and watches every channel. When it sees something going wrong, it decides within milliseconds what kind of fault it is and how urgent it is. It then either takes a low-consequence action itself or raises the alarm and asks the control system to act. Seconds later a person has a written work order.
Signals are read through read-only connectors, turned into compact fingerprints and judged against known faults and fixed rules. The Decide budget, from the signal arriving to an action or request being written, is under 30 ms (target, p99).
The Unovie Reasoning Model looks at the event, the equipment's history and the site's playbook, then drafts a work order. The first line is ready within 2 seconds and the full write-up within 15 seconds (targets). The model never moves equipment.
At 02:00 the Kiln replays the day and proposes at most three small edits to the playbook. An edit is kept only if it does better on a locked set of past events. Model weights never change.
AgentMesh sits at the site operations level of the factory network (Purdue Level 3), between the machines and the business systems. It reads signals from sensors, controllers, historians and test rigs, and writes only to the outputs the site has listed for it. The safety PLC and the safety system stay in charge of anything that can hurt people or equipment. Toward IT, the pod only sends: finished work orders go out to maintenance and quality systems, and nothing comes back in on that path. Findings can also go to an AgentNexus rack in the customer's own enclave, which works across many sites.
The pod's work is split into two roles. The Reflex role holds the fast path: it owns the connection to the machines and is the only part allowed to write outputs. The Thinker role holds the slow path: reasoning, records and the nightly review. Each role runs on its own server, and each role has a warm standby on another server, so either server can carry everything if the other fails. The four-server pod has two such pairs (section 4). Everything runs as microservices on CNCF-conformant cluster software.
| Service | Stage | Role | What it does |
|---|---|---|---|
| Unovie Gateway | Sense | Reflex | The boundary to the machines. Read-only connectors (DAQ, OPC UA, Modbus, CAN, MQTT, receive-only span port), an allow-list per protocol, and an audit log. |
| Ingest Buffer | Sense | Reflex | Holds each channel's latest signal in memory and cuts it into windows. |
| Signal Encoder | Sense | Reflex | Turns each window into a compact fingerprint of what the signal looks like. |
| Unovie Decision Engine | Decide | Reflex | Compares the fingerprint with past events, applies the site's fixed rules, and acts or raises a request. |
| Vector Memory | Decide | Reflex | In-memory index of past events: "have we seen this before, and was it bad?" Copied to the standby. |
| Reason Service | Reason | Thinker | Gathers the context for an event, asks the Reasoning Model, checks the answer and sends the work order. |
| Unovie Runtime | Reason | Thinker | Serves the Unovie Reasoning Model locally, with its weights pinned by checksum. |
| Data Triad | Reason | Thinker | Three stores: an analytics store for signals and history, a knowledge graph of equipment and rules, and the Vector Memory archive. |
| Kiln | Kiln | Thinker | The nightly review that proposes, tests and keeps or drops playbook edits. |
| Mesh Arbiter | All | Both | Watches the other server, decides who owns each role, and moves a role when a server fails. |
| Evidence Log | All | Both | An append-only, signed record of every decision, action, request and playbook change. |
The Sovereign Pod is one 3U shelf holding four compact servers, all the same model: AMD Ryzen AI Max+ 395, with processor, graphics and AI accelerator on one chip. Unovie custom-builds the pod to meet agentic performance, low-latency and high-availability requirements, and delivers it configured and tested. Because every server is the same, there is one spare-parts kit and one software image, and a replacement server sets itself up from the slot it's put in.
The four servers form two pairs. Each pair has a Reflex server and a Thinker server that stand in for each other, as described above. A pair can watch its own line or its own set of domain kits. A pod can start as one pair and grow to two without a rebuild.
| Item | What it is | What it's used for |
|---|---|---|
| Processor | AMD Ryzen AI Max+ 395, 16 cores | Gateway, rules, Reason service, the system itself |
| Graphics processor | Radeon 8060S, on the same chip | Signal fingerprints and the decision models on Reflex servers; the Reasoning Model on Thinker servers |
| AI accelerator (NPU) | XDNA 2, about 50 TOPS | Reserved for small helper models once they're tested on it |
| Memory | 64 GB unified, shared by processor and graphics (128 GB on server 4) | Models, the in-memory Vector Memory index and signal buffers. Memory is what we plan around most carefully. |
| Storage | NVMe solid-state drives in three tiers | A system drive; a fast drive for the live index and working data; a four-drive card for signal history and the Data Triad's archive |
| Network | Two 10 Gb/s ports, plus direct cables to the other three servers | One port to the machine network and one to the site network, so the two never share a cable |
| Choice | Why |
|---|---|
| Two Reflex/Thinker pairs | Each pair can lose a server and keep deciding. Two pairs double the lines or kits one pod can carry. |
| Cluster control on three servers | Cluster control needs a majority to keep working. Three members survive losing any one; the fourth server gives all its memory to work. |
| More memory on server 4 | Room for longer reasoning context, more domain kits and the nightly Kiln. Pair B's essential work still fits a 64 GB server, so if server 4 fails, server 3 carries it and only the extras pause. |
| Direct cables in a full mesh | Every server is one hop from every other. Heartbeats and copies of data don't depend on the site switches, and a failed cable has another path. |
| Separate machine and site ports | The machine network and the business network never touch on the same port, which keeps the boundary simple to inspect. |
| One model of server | One spare-parts kit, one image, and any server can replace any other. |
An event goes through the same steps every time. There is no language model on this path: every step is either a small, fast model with a fixed set of answers or a fixed rule, so the same input always gives the same, auditable answer.
| Step | What happens | Budget (target) |
|---|---|---|
| Read | The Gateway receives the sample and puts it in that channel's buffer. | 1 ms |
| Fingerprint | When a window is complete, the Signal Encoder turns it into a fingerprint. | 8 ms |
| Recall | Vector Memory finds the nearest past events and whether they were faults. | 2 ms |
| Check rules | The Policy & Bounds Gatekeeper applies the site's fixed rules: which outputs are allowed, limits, hold-offs. Alarm text and equipment names are picked out. | 12 ms |
| Write | The Lane A action or Lane B request is written to its output. | 2 ms |
| Margin | Kept free for bursts. | 5 ms |
| Total, signal to action or request, p99 under full load | < 30 ms | |
Only for outputs the site has listed as Lane A, such as a reject gate, a conveyor divert or quarantining a device on the IT side. If Decide is unhealthy it stops issuing Lane A outputs: no output is the safe state.
Anything touching process safety. AgentMesh raises the alarm and writes a request to the PLC or safety system, which decides. If AgentMesh is silent, the PLC's own logic still runs.
A small decision model adds context: should someone inspect, and how urgent is it? It runs after the Lane B request has been written, so it can never slow down the alarm.
Every decision is handed to the Thinker with a short snippet of the signal. The Reason Service gathers what the site knows about that machine (its nameplate, recent events, active alarms and the last work order) from the Data Triad, and asks the Unovie Reasoning Model to write it up. The answer is checked against a fixed form before it goes anywhere.
A site's playbook holds what it has learned: which patterns mean what, which checks to run, how to word a work order. The Kiln improves the playbook, not the model.
| Step | What happens |
|---|---|
| Replay | At 02:00 the Kiln replays the day's events and outcomes from the Data Triad. |
| Reflect | It looks at where the playbook was wrong or slow and proposes at most three small edits. |
| Test | Each edit is scored on a locked set of past events that it never learns from. An edit that doesn't do better is dropped. |
| Keep | Edits that pass are switched on and written to the hash-chained audit trail, so every change can be traced and undone. |
AgentMesh is designed against NIST SP 800-53 controls and the NIST AI Risk Management Framework. It is not a safety function and carries no SIL rating: the PLC and the safety system remain the safety layer.
| Concern | How it's handled |
|---|---|
| Writing to the wrong thing | Connectors to the machines are read-only. The pod writes only to the outputs listed in the site's output map, and only the server that owns the Reflex role may write. |
| Data leaving the site | Nothing leaves the site except finished results the customer has chosen to send. The path to IT is outbound only. |
| Tampered models | Model weights are pinned by SHA-256 checksum and checked before loading. |
| Services talking to the wrong service | Every service has its own identity. Traffic between services is encrypted with mutual TLS, and the network policy denies anything not explicitly allowed. |
| Knowing what happened | The Evidence Log is append-only and signed. Every decision, action, request and playbook change can be traced to its inputs. |
| Losing records | Records are copied to the standby server, and the Data Triad takes a nightly snapshot that is copied off the active server. |
The Mesh Arbiter on each server sends a heartbeat every 10 milliseconds on three separate paths. If a server misses ten heartbeats in a row on all three, about 100 ms (target), the other server takes over its role.
| What fails | What happens | Who handles it |
|---|---|---|
| The Reflex server | The Thinker server, already warm, checks that nobody else holds the machine-side address, takes it, and resumes Decide. It keeps doing Reason as well. | Mesh Arbiter |
| The Thinker server | The Reflex server starts its standby Reasoning Model and intake, and keeps deciding. Work orders resume once the model is warm. | Mesh Arbiter |
| One heartbeat path | Nothing changes; the other two paths still carry the heartbeat. | Mesh Arbiter |
| A server that comes back | It stays on standby. Roles move back only when a person decides to: there is no automatic failback. | Operator |
| Decide itself is unhealthy | Lane A outputs stop. Lane B falls back to the PLC's own logic. | Decision Engine |
Every kit runs on the same loop. A kit is a set of connectors, encoders, a playbook and an equipment graph for one kind of problem. Images can confirm or locate a fault, but they are never a kit's main signal. The first kit built end to end is Rotating Equipment Health.
| Kit | Main signal | Lane |
|---|---|---|
| Rotating Equipment Health | Vibration DAQ, historian tags | B (or A: slow-down request) |
| Leak & Energy Loss | Ultrasonic DAQ, flow and pressure tags | B |
| Battery Safety Fusion | Read-only BMS, off-gas, thermal | B |
| Electrical Health | Partial-discharge couplers, power-quality meters | B |
| Test Campaign Monitor | Read-only mirror of the test-rig DAQ | B (advisory) |
| Process Golden-Run | Historian and OPC UA tags | B |
| OT Network Watch | Receive-only span port | A (IT-side quarantine) |
| Alarm Flood Triage | Control-system alarm stream | B |
| Process Signature QA | Weld and machining signatures | A (reject gate) |
| Document Copilot | Manuals, procedures, work orders | None (no outputs) |