In AI Process Simulator the agent plans the components, fluid package and equipment, you review the specs and assumptions, and the simulator's own solver computes every result. It cannot delete equipment or rewire connections yet, and five to ten unit flowsheets cost about $0.05 to $0.12 in API usage in his testing.
The 20-year deal funds uprates at 11 nuclear units. Constellation says it will build agents for interconnection planning, outage and turnaround management, and OT security. All of the AI work is planned. Outage and turnaround planning is named as agent work at a major operator, so watch what actually ships.
An AI coding agent walks raw sensor data in S3 through features, training and deployment to SageMaker or the edge. The new example uses real SCADA data from a low-pressure gas compressor that Shell released for a 2020 challenge.
For reliability engineers: go from raw sensor history to a first failure prediction model.
Compressor example by Walt Mayfield
AI Model Builder turns specs, drawings and BOMs into product models, AI Seller drafts priced quotes from RFQs, and a person reviews all output before it is used.
For industrial sales: start from an RFQ and get a draft quote that is configured, priced and tied to the BOM.
One outage costs the refinery $35 million, so the agent watches sensor data for early warning signs and hands engineers a ready maintenance case with past work orders and spare parts lead times.
AI demand is now big enough to make a utility build a new 765-kV substation and transmission lines.
The team mapped how four equipment teams diagnose trips and profiled maintenance records across more than 830,000 assets, and fewer than half of equipment types had enough usable history to start. Capture how your best engineers diagnose before you pick a model; engineers keep the final decision.
Built by a piping and pipe-supports engineer, it gives an agent 50 tools for layers, block tags, ISA instrument balloons and support registers that can be compared by tag across revisions. It is locked to one work folder by default, can keep the original drawing untouched, and logs every edit.
For piping designers and pipe-support engineers: have the agent draft, dimension and export the routine sheets, then check them like any other drafter's work.
With AVEVA owned and Cognite pending, Schneider is putting design, plant and energy data in one place. That's the data AI agents need.
Plants and data centers over 200 MW could get onto the grid faster than through the standard queue.
The screen and the agent drive one engine, reads are always allowed, and any live write to the PLC needs a one-time confirmation token. Download to hardware is left unbuilt on purpose rather than guessed.
For controls engineers running Delta PLCs: let an agent scaffold and check ladder logic, with live writes gated behind an explicit yes.
It connects to SAP PM, Maximo, UpKeep, OPC UA and MQTT. As of version 0.4.0, by default it waits for a person's OK before it creates a work order or makes any other change.
For reliability engineers: ask "what's the history on pump P-201?" and get the work orders, parts and the manual's procedure from your own systems.
The hosted server ranks contract manufacturers by process, material, certifications and quantity from a directory built from OpenStreetMap, Overture, Wikidata and company websites, and sending an RFQ needs the user's consent. A manufacturer's own agent can answer the RFQ, so quoting can run agent to agent.
For buyers and hardware founders sourcing machined or fabricated parts: get a ranked shortlist and a structured RFQ from one request.
It works on the project files directly, with no EAE running, and covers IEC 61499 function blocks, CATs, OPC UA mapping and HMI screens, including an ISA-101 style review. It does not compile or download to devices, so EAE stays the place where the build gets checked.
For automation engineers on EcoStruxure: get an explanation of an inherited project, or a first pass at a function block, before you open EAE.
Built as a free alternative to Factory I/O, it connects over OPC UA, Modbus TCP, S7 and PLCSIM, and a grader runs your program against 18 scenes while it cuts the E-stop and injects faults. A broken wire on the pressure transmitter reads 32767, so your logic has to tell a fault from a reading.
For automation students and instructors: the first-hour guide goes from download to OpenPLC driving a sorting line, with no Siemens licence needed.
Conpot pretends to be industrial gear and logs whoever tries to use it, and its 1.0 release on Sep 20 rewrote the platform after the last release in 2018. A decoy placed where nothing legitimate should connect turns any connection into a lead worth chasing.
For OT security and plant network teams: a cheap way to learn who is probing control system ports before agents and remote tools add more of them.
Found via Justin Middler's write-up
Load an XER file or connect to P6 EPPM and ask for critical path, float, earned value and baseline changes. Write-back is off by default, which is the right setting for a schedule an owner will audit.
For project controls schedulers: run the DCMA checks in minutes instead of days, then decide what to fix yourself.
Point it at a well database, set decline parameters, and get a valuation and a map in chat. Its own README says Claude plays reservoir engineer and the server is the calculator, so the decline curve is still a judgment call.
For A&D analysts and reservoir engineers: get a first-pass valuation in chat, then check the curve yourself.
Process engineer Scott Duncan packaged plant practice into a skill, so Claude checks its PFD against standards and a review checklist before handing it over, catching errors like an open control loop or a steam coil above the liquid level. The output is an editable SVG plus the script that made it.
For process and chemical engineers: a first draft of a conceptual PFD in minutes, to mark up like any junior engineer's work.