An AI agent programmed a Siemens S7-1500 through an MCP bridge
The useful part is the refusal: the agent knew the E-stop belongs to hardware safety, not code.
The useful part is the refusal: the agent knew the E-stop belongs to hardware safety, not code.
Ahmed Abdel Rahman splits geomodel quality checks into six agents, each with one job, such as finding gaps in well logs, checking upscaling and comparing 3D properties with well data. The agents point to where an expert should look and leave the modelling tools and the geomodeller's judgment in place.
Claude writes the process route with stations, cycle times, breakdown rates and yields, sizes each station from demand, and a discrete event simulation computes every number on screen, from throughput to the bottleneck.
For manufacturing and industrial engineers sketching a new line: click a station, change its cycle time or buffer, and watch where the bottleneck moves.
Davy Demeyer gives Claude the batch size, batches per day and one reactor tank, points it at the Acceleer MCP server, and lets it run for a couple of hours before a first design review. The engineer's job moves to the review; he is taking a limited number of paid proofs of concept for chemical, pharma and food plants.
The costly mistake is usually not a bad meter but a flow computer running on a stale gas composition. Their line is that an unvalidated model output should never change a billed quantity.
An agent skill and MCP toolkit that builds native SLDPRT and SLDDRW files, exports a PDF and rereads dimensions. Its stated rule is that a tool call succeeding is not the same as the task being done. The README is mostly Chinese.
For mechanical engineers: a worked path from a plain request to a checked part and drawing, with evidence kept, which matters more than speed when the part goes to a shop.
Open-weight models run on site so plant data stays local. In the author's own tests on a benchmark built from IBM AssetOpsBench data, a 4B model passed 142 of 143 tasks on a 4-CPU machine, and an approval gate stopped every injected write in its attack tests, though planted false values still got through.
For maintenance and reliability teams: evidence that a cheap local model can handle work order questions, and that a human approval step is what protects the records.
The post argues that the same "what if" models once limited to companies like Boeing are opening up so a small shop can stress a design, and a student can explore a turbopump, before metal is cut.
Firdauzi describes a stack that traces why gas is flared, forecasts monthly volumes and reads flame imagery, aimed at upstream operations that still treat flaring as a monthly surprise.

Scott Duncan builds an assistant on real data from a coal-fired boiler. Through one MCP server, it connects Claude to a historian, the DCS alarm logs, an asset knowledge graph and the plant documentation.
For operations and reliability engineers: a step-by-step pattern for an agent that reads your own historian and alarms before it answers.

An open source plugin that turns text or images into STEP, STL, GLB and 3MF files from inside a coding agent.
On Oct. 9, FERC turned down the core of PJM's plan, including interim service and bring-your-own-new-capacity rules, saying those parts reach into state control of retail rates and appear to discriminate against large loads. It did accept a definition of large loads and a registry of them.
Plans, casing designs and pressure profiles flow from Oliasoft into Corva, and as-drilled data and NPT events flow back to tune the next well design. Customer pilots are still in preparation. The next well plan is meant to learn from the last well.
Under a framework agreement, new explorer Galdor gets TGS data and machine learning workflows to find prospects on the Norwegian shelf that older methods may have missed. A small explorer can mine old seismic for prospects others missed.
The start-up's top candidate is meant to make ammonia more efficiently through the Haber-Bosch process, and it has raised $10 million. A better catalyst for a high-volume chemical could cut costs at existing plants.
In one 2026 study, only 19% to 37% of AI-generated HAZOP scenarios held up. Use AI for defined checks against existing data, decide up front what it may and may not do, and keep the PHA team in charge of conclusions. A PHA is a team's judgment on how a process can go wrong, not just a document.
It predicts UV correction factors from a molecule's structure. XtalPi reports a 4.74% mean yield error across 178 reactions, a vendor figure, and results depend on the compounds and analytical conditions.
For process chemists: estimate yields from LC-UV runs without making a pure reference sample first.
An open source takeoff tool with a built-in MCP server, so an agent like Claude Code can measure rooms, count symbols and work out materials and waste.
Describe a machine in one line and the skill has Claude build an interactive isometric wireframe you can rotate and inspect, with WebGL effects for fire, water and plasma and individual parts that move. It installs as an agent skill and is open source.
For sales and training teams: a quick way to mock up how a piece of equipment works for a pitch or onboarding. The post shows nothing about engineering accuracy, so treat it as illustration, not a drawing; for ops, upstream and chemical engineers the fit is a stretch.
Through the TE1000 Automation Interface it can walk the System Manager tree, author PLC code, configure EtherCAT IO, link variables, build and deploy, with runtime actions confirmation-gated and off by default.
For controls: use an MCP client beside a running XAE Shell to draft POUs and IO maps the way an engineer would in the GUI, without giving the agent unsupervised download rights.
It checks a view against the Perspective schemas on your Ignition install, deploys it to a sandbox project on a development gateway, and opens it in a headless browser so clipped text, bad tags and script errors come back on the component. It is for development gateways only, and it is not affiliated with Inductive Automation.
For controls engineers: have an agent build a Perspective view, deploy it to a dev gateway, and read back what failed.
Hitachi's field agents are deployed, Santee Cooper's 75% faster modeling is a forecast, and Constellation's agents are planned. The digest notes none of the documents mention NERC CIP. It gives operators a quick way to tell which AI claims are already running and which are still promises.
A ready-made skill folder you can add to Claude Code, Claude.ai or the API so it works with spreadsheets the same way every time.
For industrial sales: turn a customer's spec sheet into a priced quote in Excel.
Over several runs it sorts the team's rows into ones the AI finds every time, ones it finds sometimes, and ones it never finds. The author's point is to let AI draft the generic rows and save the experts for the hazards unique to the project.
For process safety engineers and PHA leaders: see exactly which hazards an AI misses before trusting it with any part of a study.
Bechtel says design changes can now go straight to autonomous equipment in the field and be worked the same day. It is building three LNG trains there for Woodside. Design changes reach the field in a day instead of days or weeks.
Built on Equinor's open Volve dataset, it wrote its pass and fail tests down before each run, and its README reports the ones that failed. One finding: none of 3,673 drilling report events fell inside a production problem window, because no rig is on a producing well to write reports.
For production engineers: a good reminder that an agent can only explain a well problem if somebody recorded what happened.
Lead times more than doubled since 2021, so the steel and windings that connect new generation to the grid are now a binding constraint on data center power, not a side detail.

HVAC, electrical, compressed air, pipe and production equipment sit in one connected model; the point is agent-driven BIM edits for industrial buildings, not Q&A over a drawing.

A Bain partner says power, electrical labor and community opposition now gate the boom, and that a gigawatt campus looks more like a power project with a large computing load than a normal building job.
A Limble survey of 686 U.S. maintenance and operations professionals finds teams finish just 54% of scheduled preventive maintenance, and 47% of leaders report more than 10 hours of unplanned downtime in a typical month.
One agent checks photos of every tool and wire before and after a job. Hitachi reports up to 30% higher productivity, a vendor figure. The agents were built by the field crews who do the work.
Drivers cover Modbus, OPC UA, S7, EtherNet/IP, DNP3, IEC 61850 and more. An MCP gateway lets outside agents query devices with per-tool authorization.
For OT engineers: pull plant device data into one API and give an agent controlled access to it.
The authors describe AI that learns from past project schedules, drafts early schedule structures and helps planners test scenarios, with planners kept in the loop. They say the field is still early. Most companies already hold decades of old schedules that could inform the next project.
Cosmo will try new GE Vernova asset software on real refinery maintenance, including AI agents that draft FMEA templates and summarize inspection reports.
The refiner ties its AI work to earnings, uptime, maintenance costs and safety, and uses it for predictive maintenance in refining and leak detection in midstream.
The twin uses live plant data plus grid and energy market inputs to plan production and power buying, and Siemens estimates it can save up to 15% of energy.
A personal open-source project by Łukasz Czarnacki that lets an assistant search and read TIA Portal, STEP 7 and WinCC Unified docs live.
For controls engineers: ask about a PLC or HMI setting and have the assistant pull up the matching manual page.
The solid-modeling engine that does booleans, fillets and shelling is now free under Apache 2.0, and until now only legacy licensees could get it. A free Python wrapper, smlibpy, adds an MCP server, so AI agents can build parts and export STEP files. The poster works at AWV Informatik, the company behind smlibpy.
For engineers: an agent can now draft real CAD geometry on an industrial-grade CAD kernel.
When the SCADA vendor ships MCP itself, the question stops being if and becomes who approves writes.
For Controls and SCADA engineers:
Version 2.7.4 drives TIA Portal V20 or V21 to add hardware, generate tags, DBs and SCL or LAD blocks, and compile. It is free under MIT and is not made by or affiliated with Siemens.
For controls engineers: generate PLC blocks and put a TIA Portal project under Git from one prompt.
The German start-up's software pulls values from PDFs, CAD files and emails with a citation for each, flags conflicts for an engineer to approve, and drafts RFQs. Linking design quantities to tender prices is still planned. Owners of storage and substation projects could spend less time retyping data between design and procurement.
It is research code from a Seoul National University lab. It needs Windows and a licensed copy of Aspen Plus, and it was built to test local open models on process simulation.
For chemical engineers: let a model set blocks and read stream data without writing raw COM code.
This free, open-source tool has about 2,800 stars on GitHub. It lets an AI assistant create and edit models and run FEM analysis inside FreeCAD.
For mechanical engineers: describe a bracket, flange or spool piece in plain words, then run a quick stress check before it goes to drafting.
The tools turn smart meters utilities already own into grid sensors and use anomaly detection to flag early warning signs before customers lose power.
The OPC Foundation shipped it in release 2.0.0 of its .NET stack. It lets an AI assistant browse, read and write OPC UA servers, with smaller profiles to limit what it can touch.
For controls engineers: have an assistant browse a server, read live values and pull tag history.
Arsenal-2 opens in Baltimore in 2030, with one system running fabrication, inspection and paperwork from day one.

Voice is how predictive maintenance reaches the operator on the pad, not the dashboard.
For Midstream reliability and compression teams:
An agent can write function blocks, compile and fix errors, rename variables across a project, run a simulation and download to a PLC, picking the right CODESYS version from the project file. Whether a download or PLC write needs a person's approval is set in the AI client, not the server, so that setting is the one to get right.
For controls engineers on CODESYS-based PLCs: hand off the routine edit, compile and release chores while keeping each tool call visible.
Nobody had told the system which assets were critical, on standby or already down for a planned shutdown, so it flagged everything with equal confidence. Pratik Dungarwal's point: the path runs through a clean asset hierarchy, criticality ranking and failure mode libraries first.