Leads at Pella and Hexion describe pilots that free time for problem-solving, while Hexion's manufacturing IT lead warns that AI vendors often do not understand chemical plant safety and compliance.
Helenio Gilabert describes a system where separate agents track servers, controllers, disk health, network errors and alarm floods to ISA-18.2, and one agent rolls it all into a health score per piece of equipment. A scheduler agent turns failure predictions into work orders, and everything runs air-gapped on site.
The software runs on top of the existing SCADA, and the unnamed client also reports at least 2.5% lower annual power costs and up to 4% more throughput on a line apportioned about 37% of the time. These are vendor figures; the full results are due at the AVEVA Pipeline Summit on Oct 27.
TCG Digital says HPL is using its mcube agent platform to optimize processes in real time and push yield and energy performance at India's largest phenol and acetone plant. That is a vendor claim from inside the same group with no numbers yet, so watch for results.
Titan lists at $899,000 and prints structures up to 27 feet tall, and ICON says it cuts wall cost by about 40% (company claim). The company reports more than 270 printed buildings, including 10 Army barracks under way at Fort Bliss.
Across eight scenarios with no prior knowledge of the systems, the models met every objective, and in one they found a robotic arm, mapped its limits and moved it. No single safeguard stopped them every time.
Josh Cook, a Microsoft MVP, says Apps in Copilot Studio is now in public preview. You describe what you need, refine the draft by chat, connect approved data, test, then publish and share. This is how a plant or sales team could build a small internal app without a developer.
Rajat Bhageria argues a robot that does one task at 99.9% can be worth more than one that does 100 tasks at 70%, because at 70% "you don't even get a sales call." It's a fair test for any AI tool pitched to an operation. The episode is sponsored by CoreWeave.
At SCSP's AI+ Robotics Summit, FANUC America's Michael Cicco described physical AI as a standard robot paired with a much larger AI model. He said independent safety layers matter when AI directs machines, and that robot technicians should be treated as a skilled trade.
Ilir Aliu, who NVIDIA invited, lists GTC Berlin sessions that include Mercedes-Benz on factory robotics and a panel with Sereact, Agile Robots, Wandelbots and Dexory on European robotics, where he wants to hear what they actually deploy.
An MCP server that checks a PLC's online state, snapshots data block values and captures monitored watch tables, so nobody has to take screenshots by hand. It's early, tested against a fake TIA Portal and once against a real PLC.
For controls engineers: let an assistant see what the PLC is actually doing while you debug.
The new map layers help assess grading challenges, anticipate stormwater infrastructure needs and spot slope limits on buildable area. Cedar is an Austin and San Diego housing design firm and says cedarOS also pulls zoning and environmental data.
For construction and capital projects teams: check site grading and drainage constraints before committing to a design.
Microsoft says Today is a team knowledge base curated by Copilot and agents and powered by SharePoint, meant to fix fragmented and outdated information so agent teammates have the context to perform tasks and answer questions.
Elie AlChaer ran change-order line items through two models and reports both got every item right, a small but concrete preconstruction check for capital project teams.
The company says it resolves customer calls, responds in under a second and costs less than other frontier models. Those are the company's own claims, with no benchmark in the post.
The server runs spectral analysis, bearing fault matching and ISO 20816-3 severity locally, refuses to guess from filenames, and ships a Claude Code plugin with eight diagnostic skills.
For reliability: load a vibration CSV, ask whether a bearing is healthy, and get matched fault frequencies plus an HTML report while the raw signal stays on your machine.
Check your MCP server, get a quality score out of 100 and see what to improve. Manufact is offering a free month on a standard plan for posting your score.
For controls teams: several industrial MCP servers are now in this feed, so a scorecard is a cheap check if you build or buy one. It is developer-facing and a promotion, so the fit is a stretch.
Kenny Shultz says he got partial MEP drawings in minutes that would take at least a day by hand. Evidence is screenshots only and the time claim is his own.
For construction MEP and HVAC design teams: a prompt-to-drawing workflow worth testing on a past job before trusting it.
An assistant handles one task in seconds and is paid from the software budget. An agent handles one interaction in minutes and is paid from the labor budget. A long-running agent works toward a goal for weeks or months, so the seller can price on the outcome.
Samnang Aing posted an 8-minute video on how a flat floor plan turns into a 3D house as you scroll, and how he directed Claude Opus 5.5 with no magic prompt. The starter prompt and repo are public. The demo is a concept page for a 946 m² courtyard house.
For construction and sales teams: a possible way to show a client a plan in 3D. It is a concept page with no engineering data, so for ops, upstream or chemical work the fit is a stretch.
A new CAD app built in Rust with sketches, constraints, assemblies, sheet metal and STEP import and export, plus an MCP server for agents.
For maintenance and capital project teams: model brackets, enclosures and replacement parts at no license cost; the project calls itself early but usable for simple parts.
PJM filed a study agreement for SoftBank's Flagship project on federal land in Pike County, Ohio. It covers an added 9.6 GW of data center load and 10.2 GW of new gas generation, on top of an initial 800 MW phase, and SB Energy intends to pay for that plant and the lines.
Agents classify building alarms and read live HVAC data for equipment trouble. At NVIDIA's Voyager building in Santa Clara, Honeywell's own pilot figures show up to a 90% average drop in unnecessary alerts and up to a 10% potential cut in energy use.
Applied Computing's Orbital model flags equipment and process issues, and Seeq gives teams the tools to investigate them. KBR analysts will use them first, writing reports for ammonia, petrochemical and LNG customers.
It gives Claude, Codex and other agents 28 tools that run through Revu's own script engine, and when a license tier blocks headless scripting it does the job with open source PDF libraries and says which path ran. It is unofficial and free only for noncommercial use.
For estimators, project engineers and document controllers: hand off the repetitive submittal and drawing set chores and check the output like any markup.
It extracts equipment, lines, valves and instruments, compares revisions, flags dangling lines and points to the page region behind each answer. Uncertain extractions go to a human review queue.
For plant and piping engineers: see how revision-aware drawing search differs from a PDF chatbot.
An agent can inspect drawings, manage layers, add entities and plot, through live AutoCAD or headless accoreconsole when COM is not available, and it only opens paths inside folders you allow. It is one of 47 host adapters that share the same MCP setup.
For CAD managers and designers: let an agent batch edit layers or plot sets without opening AutoCAD.
Most maintenance records say which part got swapped, not why it failed. Here a technician verifies each cause before an AI agent can use it.
For reliability engineers and maintenance planners: link repeat failures and bad actors to their causes, and let agents query only facts a person checked.
Google says Gemini 4 Argon can control robots, build physics-calibrated simulation twins from camera feeds or sketches, and practice a task in simulation before doing it for real. It scores 87% on the ERQA physical reasoning benchmark. For automation teams, the useful part is practicing in a digital twin first, though most of the demos are still in simulation.
IACF's server and Claude Code skills turn a design into an importable, validated Apriso .gpm package, checked against 65+ rules before it builds. It is a vendor product, access needs IACF approval, and the demo output shown is a plant manager dashboard.
For operations and IT teams: an MES project spends much of its time turning requirements into screens, views and queries, and this aims an agent at that work. Ask IACF what the approval involves before planning around it.
A hands-on workshop that builds anomaly detection, fault diagnosis, repair planning, scheduling and parts-ordering agents for a tire plant example, using Microsoft Foundry and Azure. The repo dates from January and was updated Oct 13, and it runs on services Microsoft sells.
For reliability and maintenance teams: it maps the path from sensor alert to repair plan to parts order, and the scenario fits most plants with machines and a spare parts store.
An MCP server that gives an agent project context, CPU-specific instructions and manual evidence, then saves versioned programs with GX Works2 import files. Ladder candidates come through MCP, and Structured Text generation sits in its optional workbench.
For controls engineers: draft and review Mitsubishi PLC logic with the manual's evidence attached, then import it for a person to check.
Gov. Tina Kotek told the state environmental agency to pause air and water pollution permits for new or expanding data centers, and to stop processing city wastewater changes made to serve them. OPB says this effectively pauses most large data center development in the state. Amazon says it could cost thousands of jobs.
Teams upload robot data, fine-tune a pi model on hosted compute and run it through a hosted gateway. Early partners include an autonomous wheel loader loading a dump truck in Utah. Access is by waitlist.
Cleanview's count puts behind-the-meter generation at more than a quarter of planned data center capacity, with only about 2 GW running today and Texas leading the queue.
The Claude Code plugin analyzes a legacy codebase, writes a plan you approve, rebuilds or upgrades the code, then runs old and new on the same inputs and reports PROVEN, PARTLY PROVEN or NOT PROVEN. The legacy source is left untouched and a person signs off at set points. The README says to start with one module as a pilot.
For controls and plant IT teams: this targets software codebases (the post and README name no industrial languages), so it fits old in-house tools such as reporting or data-handling code, not PLC logic. Pilot on one small module.
One suite covers Modbus, OPC UA, PROFINET, EtherNet/IP, EtherCAT, PROFIBUS, S7, BACnet, DNP3 and MQTT/Sparkplug, each with a mock device so you can test agent flows before touching a real PLC.
For controls: point an MCP client at the protocol you already run in the field, start with the mock, then flip to the live host when the prompts are safe.
Given 11 demo drawings and site conditions, the AI produced a 61-week plan, crane placement and a 6-day-per-floor cycle. It also found a 25 cm foundation depth mismatch between drawings. The author calls it a starting draft for a construction plan.
With a licensed copy of Logix Designer and its SDK, an agent can propose an online rung edit while you watch. By default it only proposes, and you decide when to test and assemble. The project says this is verified on FactoryTalk Logix Echo, not on a physical controller.
For controls engineers: have an agent propose a rung edit in Studio 5000, then test and assemble it yourself.
OMRON's filing says it agreed on Oct. 10, Japan time, to buy OnLogic and its parent, Interlogic, for $675 million in cash. OnLogic keeps its brand and runs as its own business. A close is planned for around January 2027, if regulators approve.
The petition asks utilities to report every AI system they use, what it does and how much human oversight it gets. That covers storm response, asset management and outside vendors.
The pilot has photographed about 1,000 poles in Windsor, and the results are being checked against past inspections. The COO says it won't replace human inspectors.
Use a finished project of your own with bad scans and mixed scales, set reference quantities first, ask the vendor to show what the tool missed, and time the full review, not just processing. The author works for BuildVision AI, a takeoff vendor. A wrong quantity that looks right can carry straight into a bid.
Release 3.25.0 of the open-source library lets an assistant run flashes, phase envelopes and process simulations, and each answer lists the equation of state and assumptions used.
For process engineers: ask for a dew point or a compressor case and get a computed answer, not a guess.