Nexus Intelligence

Nexus reads every layer of the factory stack through a library of connectors and writes back into the customer's maintenance system by filing work orders. The write is named with its artifact: agents can "file a preventive work order on a wear item before it ever causes downtime". A work order is a record created in the customer's maintenance management system, and the platform's own interface shows the completed action, a preventative work order submitted with an identifier. That is authenticated action in a system Nexus does not own.

The read side is unusually deep: "Nexus has a library of data connectors that natively integrate with all layers of the factory stack", "from proprietary control system file formats and SDKs to UNS." The named layers are PLC/DCS, SCADA, historian, MES and UNS, and the earlier site also named ERP and CMMS. A library of connectors, rather than one protocol, is what separates an integration surface from a data feed.

Control systems are proprietary, versioned and specific to each vendor, which is the work that makes industrial integration expensive, and the claim of going live in under two weeks is a claim about that work. Depth in one buyer's stack is not narrowness. No connector catalog, named CMMS or MES product, or authentication model is published; what stands is a documented write and an enumerated set of system layers.

Sourcegetnexus.ai section, stage 04, architecture diagramread 2026-09-14

Swarms of specialized agents go wherever the plant is losing output, each running a long investigation. The vendor is explicit that they work as a team: "Agent swarms flow to areas of underperformance, just as your process and manufacturing engineers do today. Agents work together, with each one focusing on a long-horizon task across equipment, process, quality", "to augment and enable your team." These are specialists working together on one problem, sent dynamically to where the loss is.

The investigation itself runs many steps: "An agent will trace an hourly throughput miss all the way down to the sticky valve that caused it", "whether that investigation takes minutes or hours." That is a chain of inference across equipment, process and quality data with no fixed script. The interface shows an agent investigating cap torque drift, correlating reject spikes to filler valve lag and pulling thirty-day OEE history as concurrent threads.

The whole product is published as a five stage pipeline, Map, Instrument, Dispatch, Act and Learn, with dispatch and action as distinct stages, so coordination is the platform's architecture rather than a feature of one agent. No orchestration runtime or workflow graph is named, and there is no customer authoring surface or versioning.

Sourcegetnexus.ai stages 01 to 05, hero and interface sectionsread 2026-09-14

A live model of the plant keeps growing as a context graph, so new knowledge enters through data and actions rather than retraining. Stage one builds the structure: "Nexus builds a live model of your factory (equipment, sources, sinks, and material flow), mapped to your asset inventory of PLCs and low-level equipment."

Stage five maintains it: "All actions and data compound into a context graph of your plant", "institutional knowledge that never walks out the door." Nothing is retrained to admit a new asset, recipe or work order. The sources are indexed, refreshed continuously and permissioned: the data fabric spans PLC/DCS, SCADA, historians, MES and UNS, and "agents start read-only and expand only when you approve the boundaries."

The structure is queried, not recalled: an agent traces an hourly throughput miss down to the sticky valve that caused it, a traversal of the graph rather than a model remembering a fact.

The graph holds equipment, process and quality data, recipes, work orders and maintenance, and people and downtime, so its subject is this plant's own expertise.

On citation, agents generate tickets and notifications with clear explainable rationales, and the interface shows the working thread behind a recommendation, which shows which evidence shaped an output, though less precisely than citing documents would.

Sourcegetnexus.ai stages 01, 02 and 05, trust-and-security section, architecture diagramread 2026-09-14

Every agent finding and recommended action comes to the team for approval before it takes effect, and agents start read-only. The gate is stage four of five and is named after it: "Act. You approve. Actions stick." The body reads "When agents conclude an investigation, they bring findings and recommended actions directly to your team."

A person stands between the agent's conclusion and its effect, as a published step in the workflow. The hero sentence says the same: "Nexus deploys long-horizon AI agents that monitor your line, trace every loss to its root cause, and act with your team's approval."

The customer controls how far permissions go: "Agents start read-only and expand only when you approve the boundaries", and "your team sees what each agent does, what it can access, and which actions need human approval." Access is denied by default and widened only with the buyer's approval, and the buyer sets which actions need sign off, so this is a review and approve surface rather than a pause or override after the fact.

The one place autonomy widens is scoped and still permissioned: "As they learn over time, they can begin to resolve small problems proactively", "with permission", "like filing a preventive work order on a wear item before it ever causes downtime." The vendor writes the permission clause into the sentence describing its most autonomous behavior. No approval queue interface, escalation matrix or configurable threshold is shown, so the gate rests on a named workflow stage and the vendor's repeated commitment.

Sourcegetnexus.ai hero, stage 04 and trust-and-security sectionsread 2026-09-14

The customer controls each agent's access boundary, which starts read-only, and SOC 2 Type II certification is stated in body copy. The home page's Trust and Security section reads "Security and control start at the plant floor. SOC 2 Type II certified". Type II tests how controls operate over a period rather than their design at a point.

Access runs on a permission model the customer controls: "Your team sees what each agent does, what it can access, and which actions need human approval. Agents start read-only and expand only when you approve the boundaries." The buyer decides what each agent may reach, and nothing widens without their say. The earlier site also described permissions managed by project and by user.

An unrelated platform also called Nexus publishes different certifications, which do not apply to Nexus Intelligence, Inc. There is no trust center or security page and no report offered on request, the certificate's number, auditor and period are not given, and neither single sign on nor SAML is mentioned; the footer carries only Terms of Use and Privacy Policy.

Sourcegetnexus.ai trust-and-security section and footer, read in fullread 2026-09-14

Audit logs and observability at the level of each agent come by default, so a team can see what each agent did and what it can reach. The trust section reads "Audit logs, observability, and approval controls come out of the box": "your team sees what each agent does, what it can access, and which actions need human approval." The subject is each agent, not each machine.

Each run can be followed in the product: the interface shows an agent's working thread, such as investigating cap torque drift and correlating reject spikes to filler valve lag, alongside the recommendation it produced, its severity and where it was sent. Intermediate steps are displayed and the output is attributed, so a run can be reconstructed.

The outputs carry their reasoning too, since agents generate tickets and notifications with clear explainable rationales. Continuous monitoring of controllers, historians and maintenance data watches the customer's factory rather than the agents. No retention period, export path or log schema is documented, nor any SIEM integration, and "out of the box" is the vendor's phrase rather than a specification.

Sourcegetnexus.ai trust-and-security section, stages 03 and 04, interface detailread 2026-09-14

Agents hold state across investigations lasting minutes or hours, but nothing stores what an agent concluded on one run for it to consult on the next. The vendor sells "long-horizon AI agents" and describes an investigation that runs "whether that investigation takes minutes or hours", with the interface showing a persisted agent in a working state carrying a named standing task. That is state held within a task, by an agent through a long investigation or by a watchdog between events.

Agents write into the context graph: "All actions and data compound into a context graph of your plant", "institutional knowledge that never walks out the door." But the graph holds domain knowledge about the factory, from equipment and recipes to work orders and people, whoever contributed it, rather than an agent's own memory.

No memory scope, lifetime or purge path is described, and nothing stores an agent's prior conclusions apart from the plant graph. "Every fix makes the next one faster" describes learning absorbed into the product, and agents that "begin to resolve small problems proactively" are being given wider permission, not retained state. Part of the evidence for a persisted agent is an interface label in a product diagram rather than body copy.

Sourcegetnexus.ai hero, stages 03 to 05, interface detailread 2026-09-14

Three named deployment targets are offered as a menu: cloud, the customer's own enterprise VPC, or on premises. The home page carries a Flexible Deployment Options block reading "Deploy Nexus in the cloud or on-prem", and names three targets in a row beneath the Nexus Platform heading: Cloud, Enterprise VPC, On-Prem. Two of those are customer environments, the buyer's own VPC and the buyer's own premises, presented as a choice the buyer makes.

On premises is the option that matters in this market, and it fits the product: the platform reads control systems, SCADA, historians, MES and UNS on a plant network, and industrial buyers commonly refuse to route control system data off site. The published go-live time corroborates it: "most deployments requiring less than two weeks to go-live", with Forward Deployed Engineers working on site. There is no region list for the cloud option and no installation or sizing documentation, and nothing claims an air gapped option, which is a different commitment from on premises.

Sourcegetnexus.ai flexible-deployment section and sectionread 2026-09-14

Typed agents are described but not cataloged, and each deployment is built with the customer. Agents each focus on a long-horizon task "across equipment, process, quality", dispatched as swarms to areas of underperformance, and the interface names a standing one, a preventive maintenance watchdog. These are named kinds of agent doing distinguishable work, with no catalog to choose from.

There is nothing to browse or select: the site has four items in its navigation, from Home to Book a Demo, and no product menu, agent roster or pages for individual agents. The delivery model points away from a pack: "Fully customized to your plant. Our team of Forward Deployed Engineers work with your manufacturing and process experts to build a self-improving system around your unique domain."

Agents built with the buyer by deployed engineers are bespoke work rather than prebuilt units taken from a shelf. Earlier materials described prebuilt ops agents for factory workflows and industrial coding agents tuned to output control languages, plus drag and drop workflow building; the coding agents no longer appear on the site.

Sourcegetnexus.ai stage 03, forward-deployed-engineer section, site navigationread 2026-09-14

The production line is watched continuously, and agents are dispatched the moment a loss emerges. Stage two of the platform flow, Instrument, reads "Nexus watches your line in real time, surfacing bottlenecks, WIP accumulation and OEE losses the moment they emerge", under the heading "Eyes on every loss, 24/7". Agents also fire on incidents and quality drifts, and on production constraint issues. A line watched without pause that fires agents the moment something emerges is a genuine event trigger, and a well evidenced one.

Events on the line are the only trigger: no schedule or webhook is documented, and no chat, email, voice or other conversational path, so work reaches the agents from the line rather than from people or other systems.

Sourcegetnexus.ai platform flow, stage 02, and the June basisread 2026-09-15

The customer chooses the model and supplies its own key. The home page reads "Deploy Nexus in the cloud or on-prem, with your choice of model available through BYOK." Bringing your own key goes further than a selector the vendor runs, because the model relationship, the account and the commercial terms all belong to the buyer, and it is stated as a platform property beside the deployment options.

No provider is named, which here reads as agnostic by design rather than as silence, since it sits beside an explicit offer to bring your own key. There is no list of supported models, routing policy or selector interface. An earlier product tuned large language models to output IEC-61131-3 Ladder Logic and Structured Text for industrial control; that was the vendor's own model work, not a customer choice.

Sourcegetnexus.ai flexible-deployment sectionread 2026-09-14

No API, SDK or MCP server is documented that would let an outside system call the platform. The home page does mention SDKs, but they are other vendors': "Nexus has a library of data connectors that natively integrate with all layers of the factory stack", "from proprietary control system file formats and SDKs to UNS." Those are the control vendors' SDKs, which Nexus consumes to reach a PLC or a historian, the opposite direction from a developer surface of its own.

Nothing on the site points inward. The navigation has four items, from Home to Book a Demo, and the footer carries only Terms of Use, Privacy Policy and LinkedIn. There is no developers entry or documentation host, no API reference or webhook, and no embeddable component. For a platform selling to system integrators and equipment OEMs, an API may well exist under contract, but nothing published asserts one. An earlier product line imported, edited and exported industrial project files such as Studio 5000 formats; it no longer appears on the site.

Sourcegetnexus.ai section, navigation and footer read in fullread 2026-09-14

Nothing tests or measures the agents; what Nexus simulates and improves is the factory. No change to an agent goes through a harness, scored cases or a golden set, and no accuracy figure, judge or release gate is published. The closest feature tests something else: "The live process model runs continuous what-if simulations to size every opportunity, so you always know which problem to tackle first." That simulates the plant to rank interventions; the subject is the factory, not the software.

"Self-improving" is the product's name for itself, "the self-improving OS for manufacturing", and stage five says "every fix makes the next one faster", but what improves is the factory: as throughput, quality and capacity improve month over month, the buyer builds a self-improving factory.

Nothing published measures the quality of agent output either, whether as an acceptance rate for recommendations or a false positive figure; the one quantitative claim is a go-live time of under two weeks, an implementation metric. An earlier controls engineering product validated generated PLC code against safety and regulatory requirements; it no longer appears on the site, and validating the customer's control code is the product's job rather than a test of the agent. With Forward Deployed Engineers on each deployment, a validation practice may exist without being published.

Sourcegetnexus.ai stages 02 and 05, hero and titleread 2026-09-14

Action runs through industrial system integrations, not through a browser or a computer. No hosted or local browser, desktop session or remote computer control appears. The path is machine to machine at both ends: connectors reading control systems, SCADA, historians, MES and UNS at one end, and a work order filed into a maintenance system at the other, with no screen in the loop. This is an industrial agent driven by protocols rather than one that works through engineering software interfaces.

UNS topics, historian tags and PLC file formats have no screen to break when an interface changes; what can fail here is a connector or a tag mapping, not a moved button. The agent interface is a surface people read, and the Forward Deployed Engineers who configure the system are people operating software. An earlier ability to import and edit Studio 5000 project files was file manipulation, not operating the engineering IDE. The architecture is described precisely enough at both ends that a path that drives screens is unlikely.

Sourcegetnexus.ai section, stages 02 to 04, architecture diagramread 2026-09-14