Integrating Data-Driven Insights with Domain Expertise Using Agentic Conversational Analytics for Well Completions Optimization

The design of completions and fracturing strategies present challenges that hinder the effectiveness of unconventional resources development, including a wide range of parameters, the non-linear nature of certain types of formations, and the balance between production and economic viability. Multivariate data analytics is key in evaluating those resources. This paper aims to present a workflow that leverages a multi-agent conversational system to integrate data, analytics, and domain expertise for improved completion strategies. Well and completions data from the Fayetteville Shale are integrated into the chat application. The framework features a router, code executor, and answer compiler as auxiliary agents that manage the conversation and perform computations. Additionally, there are four core agents: a data analytics agent for free-form analytics; a completions optimization agent for optimization and sensitivity studies; a model evaluator to provide information on the trained machine learning model; and a shale expert, responsible for analysis and recommendations. When a user submits a query, the system updates it with short-term memory to retain context before prompting the router. The router triggers the most relevant agents to process the request. After completing the task, a response is compiled and stored in memory to maintain the conversation continuity. The multiagent system is hosted in a chat window that relays natural language messages to a large language model. Domain expertise is encoded in the core agents through prompts and doc-strings embedded in the function calls. We demonstrate the agentic framework in action through two use cases applied to a completions dataset from the Fayetteville Shale. First, we show how the operator can perform analysis and recommendations based on queries, aggregations and descriptive statistics on completions properties for wells from a selected area of investigation. This process enables quick screening of correlations present in the data based on geographical distribution. The agent-generated analytics is displayed through responses and plots that allow recognizing patterns in completions design and well performance. The second use case illustrates the completions optimization agent being used to provide recommendations on optimal completions design and prediction of gas production per foot of completed lateral length. A multivariate regression model is used to characterize the relationship between completed lateral length, fluid and proppant loads. An optimization algorithm is then combined with the machine learning model to provide optimal parameters for maximizing gas production or minimizing costs. This paper introduces a composable multiagent framework that systematically merges domain expertise, data analytics, and optimization methodologies into a single platform. Messages passed to the multiagent system create a unique investigative narrative, allowing users to build a story through each subsequent addition to the chat history. By enabling natural language interactions with specialized agents, it offers a context-aware tool that can enhance traditional completions workflows and broaden the scope of automated upstream decision-making.