# What is an agentic workflow in oil and gas?

> An agentic workflow in oil and gas is a multi-step piece of work, such as building an inspection work pack or checking a bid, that AI agents plan and carry out with software tools, while checks and a named person's approval sit at set points. Unlike rules-based automation, the path can change from one run to the next.

- Label: Fact
- Type: Explainer
- Published: 2026-11-15T00:16:49+09:00
- By: Jun Zhong
- Researched and drafted by: Beacon's AI newsroom
- Checked by: Jun Zhong on 2026-11-15
- Canonical: https://beacon.winda.ai/library/agentic-workflow/

## Why does it matter?

In February 2026 NIST's Center for AI Standards and Innovation (CAISI) noted that AI agents can now work on their own for hours, writing code, handling email and buying goods, and it launched a standards initiative because trust in their reliability was holding adoption back [1]. The reliability gap is measurable. In a benchmark of 175 professional tasks set in a simulated software company, the best agent tested finished 30.3% of them end to end and earned 39.3% of the available credit once partly finished tasks were counted [12]. That is the practical reason an agentic workflow for O&G engineering is built around its checkpoints, not around the agent alone.

## What makes a workflow agentic?

A workflow is agentic when an AI model, not a fixed script, decides the next step and which tool to use, and then acts on the result. CAISI describes agent systems as one or more generative models plus scaffolding software that gives them tools, and notes that larger systems may hold several sub-agents coordinated by orchestration software [2]. Its security work concentrates on systems whose actions make lasting changes outside the agent itself [2]. What an agent is, and where one may act on a plant, is on [What is an AI agent in industrial operations?](/library/industrial-ai-agent/).

Two research patterns show the mechanics. ReAct, a 2022 method, has the model alternate between reasoning about the task and calling a tool such as a search, so each observation feeds the next plan; it beat imitation and reinforcement learning baselines by 34 and 10 percentage points on two interactive benchmarks [8]. HuggingGPT, from 2023, uses one language model as a controller that plans the subtasks, picks a specialist model for each, runs them and writes up the combined result [9].

## How is it different from a chatbot, a copilot or RPA?

| Question | Chatbot or copilot | Robotic process automation (RPA) | Agentic workflow |
|---|---|---|---|
| Who sets the steps? | The person, one prompt at a time [2] | A developer, in advance; the bot copies human actions [7] | The model, within a goal and the tools it is given [2] |
| Typical work | Answers and drafts [3] | Repetitive, rules-based tasks such as data entry and reconciliation [7] | Multi-step tasks that need several tools and may need a colleague's input [12] |
| Same output for the same input? | Not necessarily; developers build randomness into language models [3] | Yes, while the rules and screens stay the same [7] | Not necessarily [3] |
| Where checks sit | The reader checks each answer [5] | In testing before the bot is deployed [7] | At checkpoints inside the run and at approval [12] |

The line between the two kinds of automation is blurring. In a 2021 survey of US federal RPA programs, 32% had added features such as machine learning, image recognition or natural language processing [7].

## What patterns do agentic workflows use?

| Pattern | How it works | What the research found |
|---|---|---|
| Reason and act loop | One agent thinks, calls a tool, reads the result and plans again [8] | Grounding in tool results reduced the hallucination and error build-up seen in reasoning alone [8] |
| Planner with specialists | A controller splits the task and hands each part to a suitable model or agent [9] | CAISI counts multi-agent orchestration as part of the agent system it studies [2] |
| Draft, critique, revise | The model reviews its own output and revises it [10] | About 20 points of average improvement across seven tasks [10], but without outside feedback models struggled to fix their own reasoning and sometimes got worse [11] |
| Agents coordinate, tools compute | Agents route the work and fill in parameters; fixed routines do the numbers [14] | In methane field campaigns, including oil and gas sites, routing and parameter extraction were 92.0% accurate and emission estimates 85.0% successful [14] |

**Analysis:**
The last two rows point the same way. A model that grades its own reasoning is a weak check, so in engineering work the check that counts should come from outside the model: a calculation routine, a rule set, or a qualified person reading the evidence. A workflow is only as strong as the check that sits between its draft and its approval, and the failure modes in [Why do general AI tools get engineering answers wrong?](/library/ai-errors-in-engineering/) apply at every step that feeds an integrity decision.

## Where do agentic workflows show up in oil and gas?

- **Document-heavy engineering.** A 2024 research copilot generated piping and instrumentation diagrams (P&IDs) from plain-language descriptions through a multi-step agentic workflow and scored better than single-prompt generation [13]. Linked engineering data, such as an [industrial knowledge graph](/library/industrial-knowledge-graph/), gives such workflows something reliable to read.
- **Inspection planning.** Covered US process plants must document every inspection and test with its date, inspector, equipment identifier, method and results [6], which is the kind of record a workflow gathers into a draft work pack.
- **Procurement and quality.** Employers must make sure that fabricated equipment, spare parts and maintenance materials suit the process they will serve [6], which is where a workflow comparing vendor data with a datasheet ends up. See [AI in bid evaluation](/library/ai-in-bid-evaluation/).
- **Compliance and emissions reporting.** The methane framework above ran mostly on local hardware, keeping sensitive site data off cloud services, and produced editable reports 95.0% of the time [14].

## What does a checked workflow look like?

**Example:**
An invented scenario: no real site, operator, vendor or product.

A workflow prepares the technical bid evaluation for a fictional produced-water separator. Five bids arrive, and the datasheet has 63 lines.

| Step | Who does it | Checkpoint |
|---|---|---|
| 1 | Extraction agent reads the five bids and fills a comparison table | Every value links to its page in the bid |
| 2 | Comparison agent marks 11 lines that differ from the datasheet | Lines with no stated value are flagged, never guessed |
| 3 | Calculation routine recomputes wall thickness for the three bids that quote it | One bid's figure fails the recheck and is flagged |
| 4 | Agent drafts clarification questions for the 12 open items | The buyer sends them; the agent cannot email vendors |
| 5 | Lead engineer reviews the table and the evidence | Only the engineer can recommend an award |

**Result:** two bids are technically acceptable, one goes back for clarification, two are rejected. The agents drafted every line; none of them approved anything.

## What governance does an agentic workflow need?

- **Defined human oversight.** NIST's AI Risk Management Framework asks that processes for human oversight be defined and assessed (MAP 3.5) and that someone be able to supersede, disengage or deactivate a system that drifts from its intended use (MANAGE 2.4) [4].
- **Guarding the reviewer against habit.** NIST's Generative AI Profile warns of automation bias, the tendency to defer too much to automated output, which can make confidently stated errors more harmful [5].
- **Least privilege for tools.** CAISI reports that organizations already apply least privilege and zero trust design to agent systems, alongside newer methods such as instruction hierarchies [2].
- **Logs and an inventory.** The December 2025 joint guidance led by the US Cybersecurity and Infrastructure Security Agency (CISA) asks operators to list every AI component, log its inputs and outputs, and route passive AI recommendations through existing change management [3].
- **No path into control systems.** The same guidance prefers moving data out of operational technology (OT) by push or broker over giving AI standing inbound access, and asks for failsafes that let an AI system fail without disrupting operations [3]. See [OT cybersecurity](/library/ot-cybersecurity/).
- **Management of change for anything that alters the process.** US process safety rules require the technical basis, safety and health impact and authorization of a change to be addressed before it is made [6].

**Analysis:**
For an operator, the useful question about any agentic workflow is not how capable the agents are but where its checkpoints sit and who signs at the end. A workflow that cannot show which record each draft rests on, which check it passed and which person approved it does not fit the signature chain that inspection, procurement and change management already rely on.

## Sources

1. [Announcing the "AI Agent Standards Initiative" for Interoperable and Secure Innovation, US National Institute of Standards and Technology (NIST), Center for AI Standards and Innovation (CAISI)](https://www.nist.gov/news-events/news/2026/02/announcing-ai-agent-standards-initiative-interoperable-and-secure), 2026-02-17
2. [Request for Information Regarding Security Considerations for Artificial Intelligence Agents, 91 FR 698 (FR Doc. 2026-00206), NIST Center for AI Standards and Innovation (CAISI), Federal Register](https://www.govinfo.gov/content/pkg/FR-2026-01-08/html/2026-00206.htm), 2026-01-08
3. [Principles for the Secure Integration of Artificial Intelligence in Operational Technology (joint guidance, TLP:CLEAR), US Cybersecurity and Infrastructure Security Agency (CISA), Australian Signals Directorate's Australian Cyber Security Centre and partner agencies](https://www.cisa.gov/sites/default/files/2026-01/joint-guidance-principles-for-the-secure-integration-of-artificial-intelligence-in-operational-technology-508cV2.pdf), 2025-12-03
4. [Artificial Intelligence Risk Management Framework (AI RMF 1.0), NIST AI 100-1, US National Institute of Standards and Technology (NIST)](https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.100-1.pdf), 2023-01-26
5. [Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile, NIST AI 600-1, US National Institute of Standards and Technology (NIST)](https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.600-1.pdf), 2024-07-26
6. [29 CFR 1910.119, Process safety management of highly hazardous chemicals, paragraphs (j)(4), (j)(6) and (l), US Occupational Safety and Health Administration (OSHA), eCFR](https://www.ecfr.gov/current/title-29/subtitle-B/chapter-XVII/part-1910/subpart-H/section-1910.119)
7. [The State of Federal RPA: An Update on the Governmentwide Impact, Deployment, and Best Practices of RPA, Version 2.0, Federal RPA Community of Practice, US General Services Administration (GSA)](https://origin-www.gsa.gov/system/files/state-of-federal-rpa-report-12-2021.pdf), 2021-12-29
8. [ReAct: Synergizing Reasoning and Acting in Language Models (S. Yao, J. Zhao, D. Yu and others), arXiv (ICLR 2023), CC BY 4.0](https://arxiv.org/abs/2210.03629), 2023-03-10
9. [HuggingGPT: Solving AI Tasks with ChatGPT and its Friends in Hugging Face (Y. Shen, K. Song, X. Tan and others), arXiv (NeurIPS 2023)](https://arxiv.org/abs/2303.17580), 2023-12-03
10. [Self-Refine: Iterative Refinement with Self-Feedback (A. Madaan, N. Tandon, P. Gupta and others), arXiv (NeurIPS 2023), CC BY 4.0](https://arxiv.org/abs/2303.17651), 2023-05-25
11. [Large Language Models Cannot Self-Correct Reasoning Yet (J. Huang, X. Chen, S. Mishra and others), arXiv (ICLR 2024), CC BY 4.0](https://arxiv.org/abs/2310.01798), 2024-03-14
12. [TheAgentCompany: Benchmarking LLM Agents on Consequential Real World Tasks (F. F. Xu, Y. Song, B. Li and others), version 3, arXiv, CC BY 4.0](https://arxiv.org/abs/2412.14161), 2025-09-10
13. [An Agentic Approach to Automatic Creation of P&ID Diagrams from Natural Language Descriptions (S. Gowaikar, S. Iyengar, S. Segal and others), arXiv preprint](https://arxiv.org/abs/2412.12898), 2024-12-17
14. [A Locally Deployable Tool-Grounded LLM Multi-agent Framework for Automating Methane Emission Analysis and Reporting (Y. Yan, Z. Zhou, X. Wang and others), arXiv preprint](https://arxiv.org/abs/2608.18473), 2026-08-19

## From Winda

In Winda Fleet, an orchestrator splits an integrity question into steps, specialist agents each do 1 step, and a checker re-checks the work independently. Each draft is checked against the engines and the standards, and a named person reviews the evidence and approves. Agents never approve their own work. [See all 204 agents](https://winda.ai/products/agents/)

Published by Winda Beacon (https://beacon.winda.ai).
