Energy and resources: the estate and where agents fit

Energy and resources companies do four kinds of work. They trade in markets, plan networks and plants, run generation and network assets, and run mines and processing plants. This page shows the systems they run, layer by layer, and where agents fit. The short answer is this. Agents read plant data through a read-only broker and draft work for people. They never write to a control or safety system. The frame is the same as on every page: layers you own, a lifecycle you run, and capabilities you add.

The estate

Two kinds of producer share most of this estate. The first is a power generator or an upstream oil and gas operator with several sites. It runs plants or wells, with a control room at each site. The second is a miner with pits, a processing plant, rail and a port. It often runs them from a remote operations centre in the city. The two differ most in the line of business and OT layer (R05), where the plant and fleet systems sit.

The table names the main systems in each of the fourteen layers. Vendor names in brackets are examples, not endorsements.

Three control systems matter most on this page. The distributed control system (DCS) runs the process. Programmable logic controllers (PLC) run machines. The safety instrumented system (SIS) shuts the plant down safely when limits are passed. It is kept separate by design.

Plant networks do not talk straight to business networks. A buffer network, often called the DMZ, sits between them. Historians copy plant data across it. Some sites use a data diode, which lets data flow one way only. Agents run on the business side or in the cloud. A read broker in the buffer network serves them plant data from the copy. That broker is how the tool gateway reaches the plant.

Vendors now ship agents for single systems: one control system, one EAM, one historian. This research found no vendor agent that joins the historian, EAM, lab, permits and production accounting for one investigation. That join is where a thin build fits. The method is on agents and packaged applications.

Agents stay on the information path

One rule binds this industry. Agents sit on the information path, never the control path. The information path is reading data, checking it and drafting work. The control path is anything that moves equipment.

Nothing an agent does writes to a DCS, PLC, SIS, protection relay or turbine control. Nothing writes to a collision avoidance or fatigue system. Nothing an agent does issues a permit or applies an isolation. A person, or a certified system, makes every physical change. Safety actuation is one of the four deterministic zones, where a model may never approve.

Three vendor claims made in 2026 cross this line. This guide does not cross it with them.

• Honeywell Experion Cognition, announced 9 June 2026. Honeywell says it will "proactively act on behalf of the operator" [vendor]. A proof of concept runs at Borouge in Ruwais. Honeywell promised it for the third quarter of 2026. This research found no confirmation that it shipped.
• Cognite with ABB, announced 21 May 2026. It adds agents to ABB's safety and alarm tools. Aker BP is the first customer. The release says agents can "trigger cross-system actions" and does not say which ones [vendor].
• DataRobot with Chevron, announced 2 June 2026. A "Safe Start" agent checks conditions for robotic inspection missions. It aims to replace the permit approval for each mission with continuous checks [vendor]. The stage is not stated.

None of the three says what its agent may not touch. A control vendor acting inside its own control system needs its own safety case. Yokogawa's reinforcement-learning control at two plants is the same kind of case. It is described on the line of business and OT page. It is not an agent on the information path.

The use cases on this estate

Four worked examples sit on this estate. Each is a proposal from this guide, built for a producer. A fifth, compliance evidence, is shared with utilities. The table shows what each one reads, where it writes and who decides.

The named deployments below sit in three areas of the work: markets, planning, and generation and network assets. Each says who runs it, whether it advises or acts, and who reported the result.

Markets

Battery bidding

Battery owners in Australia's National Electricity Market (NEM) use software that bids and rebids. No person signs each bid. Tesla's Autobidder and Fluence's Mosaic are two such bidders. Fluence names Akaysha Energy among its users (pv magazine Australia, 20 November 2024). The vendors say their bidders work inside limits the operator sets, with manual override [vendor]. No measured outcome is published. The bidders are in production.

That is managed autonomy (A4) on the autonomy contract. The bidder acts inside limits people set, without approval of each bid. Its learning is curated learning (L1) or reviewed learning (L2). In curated learning, people choose what it learns from. This guide asks for reviewed learning at A4. A bid moves money and creates a formal market record. Both are deterministic zones. A zone action makes the bidder Tier 3 (strict) on the risk profile. The control is a policy check and a log the bidder cannot change, both before the market gateway. That is enforcement outside the model.

What the regulator said. The Australian Energy Regulator (AER) issued a compliance bulletin on automated bidding on 5 August 2026. It creates no new obligations. The participant owns every bid, however it is made. Bids sent straight to the market operator without a participant check carry higher compliance risk. Participants should record the logic behind each rebid. Third-party providers may be liable too. This page could not retrieve the bulletin itself. It read the bulletin through a law firm's alert (K&L Gates, 10 August 2026).

Planning

Planning agents read documents, run studies and build evidence packs. All three cases below advise. Engineers or staff decide. Each is at delegated tasks (A2): the agent does a set task and a person decides.

Network models in hours. AgenticGrid comes from Monash University with Transgrid and Essential Energy. Its agents choose which studies to run. A conventional solver does the calculation. A network model of the Broken Hill area took about two hours, against about three weeks. Race for 2030, an energy research programme, reported the result. It is a trial.

Interconnection screening. PJM, the largest US grid operator, works with Tapestry from Alphabet. The partnership was announced in April 2025. Its agents screened 811 generation filings, about 220 gigawatts, for site control in under an hour. That is Tapestry's own count [vendor]. Each finding links to its source pages. PJM staff make the compliance call. That call is a formal queue decision, so it sits in the formal records zone.

Interconnection study data. Duke Energy runs agents with AWS on parts of its interconnection studies. The agents use Duke's own simulation software and grid models. AWS says data preparation fell from about two weeks to hours (T&D World, 18 September 2026) [vendor]. Duke says its engineers make all final engineering decisions. The stage is not stated.

Generation and network assets

Self-healing grid, the rule-based baseline. Duke Energy's self-healing grid isolates faults and restores supply on its own. Duke counts about 2.2 million customer outages avoided in 2025. It also counts 5.2 million outage hours saved. It announced both on 5 February 2026. The announcement does not mention AI. Self-healing usually means fixed switching logic. It is the rule-based autonomy regulators already accept. It is not an agent, and it needs no new control.

Outage request review. CAISO, the California grid operator, uses review agents built by OATI. OATI says they analyse more than 500 next-day outage requests in under eight minutes a night [vendor]. In July 2025 CAISO reported no results yet and called itself cautiously optimistic (Latitude Media, 28 July 2025). Operators keep final authority. It is delegated tasks (A2) with curated learning (L1). It touches no zone.

Control room assistants. Honeywell launched its Experion Operations Assistant on 19 March 2026, after pilots at Chevron and TotalEnergies. It warned operators 5 to 10 minutes before alarms, and operators acted [vendor]. It advises. Schneider Electric announced its Grid AI Assistant on 4 June 2026. It says every operational action stays subject to user review [vendor]. It names no utility. This research found no utility letting an agent operate protection or switching.

Proposals and commercial work

Commercial work is not an operations workflow, but it runs on the same estate. A generator's sales desk answers a large buyer's request for a power purchase agreement. A mining supplier answers a tender. Today an originator gathers the request, the buyer's meter data and past deals by hand. A structuring analyst prices the deal from forward curves in the trading system and a spreadsheet. Credit, legal and a risk committee then approve under delegated authority. It takes weeks to months (typical, unsourced).

An agent can help with the brief. It reads the CRM, public information about the buyer and approved summaries of past deals. It drafts an opportunity brief that keeps facts apart from guesses. It lists questions for the buyer and outlines the proposal. The originator reviews every draft. Price, credit, terms, every message to the buyer and deal booking stay with people. A price or contract change moves money, so it sits in the money zone. The agent has no write path to the trading system and no send path to the buyer.

It sits in the systems of record layer (R04) and the productivity and collaboration layer (R08). It works at delegated tasks (A2) with curated learning (L1). It is Tier 2 (enhanced) while it only drafts. Other customers' terms stay out of its reach. This research found no such deployment in energy or mining.

Evidence and limits

Most figures on this page come from the vendor or the operator. None is independently audited. This research found no independent figure showing that AI cut outage time at a named energy company.

The estate is typical practice, not a survey of one company. The vendor lists come from industry knowledge, and most are unsourced. Twenty searches found no agent from Deswik, Komatsu, Modular Mining, Sphera, Quorum or SLB Avocet for these processes. Absence in twenty searches is not proof of absence.

No agent deployment was found at BHP, Rio Tinto or Fortescue by 15 October 2026. They show machine learning, platform deals and job advertisements. Adoption is early in power too. A PwC study reported in April 2026 found 61 percent of utilities still in test or pilot.

The main guidance on AI and functional safety is ISO/IEC TR 5469. It is a technical report and is not binding. NAMUR Open Architecture describes a second channel that reads plant data without touching control. That is close to the information-path rule. This round did not check its primary text. The IEA's Energy and AI report (2025) surveys AI across energy. It does not validate the agent workflows proposed here.

Next step

Pick one worked example and score it on the risk profile. Technical readers can open the mining and resources blueprint or the energy and utilities blueprint. Electricity distribution and water continue in Utilities.