Source-linked AI summary
Engineering process plants as code in the age of agentic AI
Artur M. Schweidtmann, Davy Demeyer
TL;DR
Plant information is fragmented across tools and documents, so agents inherit the same manual cross-checking burden as engineers. The Comment proposes Plant-as-Code and agent orchestration with deterministic verification; it argues that design iterations could shrink from weeks to hours while engineers retain oversight.
Problem
Plant information is scattered across applications and formats, leaving handovers manual and making cross-domain consistency checks difficult for engineers and AI agents.
Source: §1 · pp. 1–2Method
Results
Individual design iterations that take weeks today could take hours, and early industrial pilots point to substantially faster iterations but are not construction-ready demonstrations.
Source: §3 · p. 4Takeaways & Limitations
If designs can be produced and checked much faster, exploring alternatives, value engineering and plant optimization become cheap, shifting the bottleneck to construction.
Source: §6 · p. 5Human review remains indispensable because reviewer agents are less predictable and cannot provide the judgment and accountability of experienced engineers.
Source: §5 · p. 5
Abstract
from arXiv · showhide
Agentic artificial intelligence (AI) could one day shorten process plant engineering from years to weeks. Today, however, plant information is scattered across documents and tools and must be carried between them by hand, so an AI agent placed into current workflows inherits the same silos. In this Comment, we argue for Plant-as-Code: defining the engineering of a plant as a version-controlled collection of machine-readable files that follow, where possible, existing standards such as DEXPI and IFC. On this foundation, Design-Ops extends DevOps principles to plant engineering, linking every engineering item to the information it derives from and replacing manual checking with automated tests. Agents can then coordinate routine work across disciplines while deterministic tools perform the calculations. Verification sets the pace: agents propose, tools check, and engineers review and authorize every change. We outline the risks of this shift and the research, industry and policy actions it requires.
1 Why plant engineering needs a new paradigm
Process-plant engineering information is split across documents and tools, making changes dependent on manual handoffs and cross-checks. Agents placed in these workflows inherit the same silos, so generated documents are not validated designs.
1 Why plant engineering needs a new paradigm
- Project information is spread across requirements, process diagrams, simulations, 3D layouts, equipment datasheets and control narratives, each in its own application and format.The conventional workflow coordinates these separate artefacts, where information can become outdated or conflicting (Fig. 1a).Source: §1 · p. 1, Figure 1 · p. 2
- A pump replacement must be carried manually into the P&ID, hydraulic calculation, layout, instrument list and control logic.Most process-engineering handovers remain manual, unlike discrete manufacturing, where shared digital models cut design-to-inspection cycle times by roughly three quarters.Source: §1 · p. 2
- Agents inherit these silos: a quickly drafted P&ID cannot be checked against simulation or layout without repeating cross-domain checks.Tool-specific copilots likewise see only their own tool’s data, and a generated document is not a validated design.Source: §1 · p. 2
- For existing plants, modifications cannot be assessed until the current state is reconstructed from scanned drawings and outdated records.Source: §1 · p. 2
2 Engineering process plants as code
Plant-as-Code defines a process plant through machine-readable, version-controlled files, extending software practices across coupled engineering domains. The model depends on shared semantics and traceable review, not merely storing documents in a repository.
2 Engineering process plants as code
- Plant-as-Code extends code-based engineering principles across the lifecycle of process plants, where designs span coupled domains and errors have physical consequences.It defines the plant’s engineering representation as machine-readable files under version control.Source: §2 · p. 2
- Where possible, the plant files follow DEXPI, IFC and MTP standards as the data model of a single source of truth, rather than only as end-of-phase exchange formats.The repository may use compact text formats such as YAML and generate standard exchange files when needed; vendor skids can be referenced from their own folders.Source: §2 · p. 3
- Git branches let teams explore and validate design alternatives without changing the plant of record; merges require agentic and human review.Source: §2 · p. 3
- A shared repository alone does not ensure consistency: PaC also needs stable identifiers, shared semantics, validated model transformations and clear design ownership.With appropriate tooling, provenance, dependencies and change history can be recorded in normal engineering work.Source: §2 · p. 3
3 Agentic AI as engineering orchestrator
Agents can coordinate routine engineering work across disciplines while specialized deterministic software performs calculations. Engineers set objectives and constraints, review results, and retain design responsibility as iterations could become much faster.
3 Agentic AI as engineering orchestrator
- Agents can break objectives into tasks and coordinate routine work across disciplines, while specialized deterministic software performs simulations, sizing, routing and control calculations.Tool-use architectures have demonstrated this approach in other domains such as chemistry.Source: §3 · p. 4
- Engineers set objectives, constraints and design ideas, judge results, and can review them in familiar environments such as P&IDs or virtual-reality plant models.They can question an agent’s choices, request changes or compare versions [11].Source: §3 · p. 4
- Individual design iterations that take weeks today could take hours, while the familiar engineering phases remain.Source: §3 · p. 4
- Early industrial pilots point to substantially faster design iterations, but a workflow demonstration remains far from a construction-ready design.Current demonstrations cover parts of the engineering chain, including text-to-simulation and natural-language P&ID querying [11].Source: §3 · p. 4
4 From DevOps to Design-Ops
Design-Ops adapts DevOps practices to plant engineering through versioned changes, traceable dependencies and automated verification. The analogy is limited: physical designs require management of change, accountable approval and operational feedback.
4 From DevOps to Design-Ops
- DevOps provides fast, collaborative workflows, but plant engineering remains staged because plants are procured and built once and errors have physical consequences.Source: §4 · p. 4
- Design-Ops links each downstream engineering item to its upstream source, so a changed design case immediately identifies affected items across disciplines.For example, a piping line size points to a flow rate in the heat and mass balance, which points to a design case.Source: §4 · p. 4
- Automated tests run whenever a change is proposed, replacing mostly manual checking; each data point is engineered once rather than re-entered by hand.Manual re-entry consumes substantial project time and commonly causes inconsistencies.Source: §4 · p. 4
- Software practice is a workflow template, not evidence of safety: a wrongly specified vessel may already be fabricated, so Design-Ops needs accountable approval and operational feedback.Management of change and traceable operational change requests remain necessary.Source: §4 · p. 4
5 Verification is the limiting step
The authors argue that verification, rather than agent speed, limits reliable agentic engineering. They propose layered automated checks, human authorization and benchmarks to guide any expansion of autonomy.
5 Verification is the limiting step
- Reliable agentic engineering requires a closed loop in which agents revise designs in response to verification results.For example, a failed pipe-clearance check prompts rerouting and another check.Source: §5 · p. 5
- Verification should cover syntax, cross-model consistency, physical constraints and safety, while reporting preconditions, expected outcomes and uncertainty.Cross-domain effects matter: rerouting a line changes pressure drop, potentially affecting pump duty and control-valve sizing.Source: §5 · p. 5
- Reviewer agents are not verification, and human review remains indispensable because some qualities resist formalization and engineers provide judgement and accountability.The authors propose bounded autonomy: agents propose, tools check, and engineers review and authorize changes.Source: §5 · p. 5
- Shared benchmarks should measure engineering errors, human interventions and the effort required to repair a design.The authors point to SWE-bench as a precedent for measuring software-agent progress.Source: §5 · p. 5
- Agents may take on more verification only when evidence shows they match or exceed the checks they replace and regulators and operators accept the limits.With such evidence, some modifications could be engineered in days rather than months.Source: §5 · p. 5
6 Outlook
Faster design and verification could make exploring alternatives inexpensive and shift engineering effort toward faster, cheaper construction. The authors also identify workforce, concentration and governance risks, alongside research, industry and policy needs.
6 Outlook
- Faster design and verification could make alternatives, value engineering and plant optimization inexpensive, shifting the bottleneck to construction.The authors cite modular units fabricated off site as one way to make plants faster and cheaper to build.Source: §6 · p. 5
- Capturing engineering knowledge in shared models, tools and checks could make expertise easier to transfer to regions with few experienced engineers.The authors also suggest faster engineering could shorten deployment of energy-transition processes and medicine production in a crisis.Source: §6 · p. 5
- Automating routine checking and data transfer risks depriving junior engineers of craft-building experience, while overtrust and flawed shared checks could spread errors across designs.The passage also identifies machine-readable plant models as cyberattack targets.Source: §6 · p. 5
- Dependence on a few providers could replace fragmentation with single points of failure, alongside new intellectual-property questions.Source: §6 · p. 6
- Academia needs research on shared engineering semantics, quantified verification confidence, realistic benchmarks and training engineers to supervise agents.Industry needs open agent interfaces, while policymakers and regulators need rules for liability, accountability and certification.Source: §6 · p. 6
- The central technical challenge is connecting heterogeneous engineering models into one coherent, verifiable plant description, with engineers retaining responsibility for consequential decisions.The authors position Plant-as-Code as the foundation and Design-Ops as the way to organize controlled changes.Source: §6 · p. 6
Declaration of AI use
The authors used AI tools to assist with drafting and related work, while retaining responsibility for the manuscript.
Declaration of AI use
- The authors used GPT-6 and Claude Opus 5.5 for drafting, literature searches, language editing and figure generation, while independently developing and retaining control over the scientific content.They report critically evaluating and revising AI-assisted content and verifying cited sources.Source: Declaration of AI use · p. 6
Competing interests
The competing-interests statement identifies the author's commercial connection to AI-assisted plant-engineering software.
Competing interests
- Davy Demeyer is the founder of Acceleer, which develops commercial software for AI-assisted plant engineering.Source: Competing interests · p. 6