15 min read
Oscar Bastidas Jossa, Artificial Intelligence Engineer
Alberto Calvo, CTO and co-founder
Summary
Quantum computing holds transformative potential, yet its practical adoption remains hindered by complexity and uncertainty about real-world applications. To address this gap, we present the Qcentroid multi-agent AI tool that systematically translates business needs into scientifically grounded quantum computing use cases through a three-stage architecture: Interview, Generator, and Deep Research. In the Interview stage, an interactive AI agent collects organizational objectives and computational bottlenecks, producing a structured problem summary. The Generator stage validates problem suitability and synthesizes candidate quantum applications by aligning business needs with established quantum computing algorithms. The Deep Research stage employs a parallelized, multi-agent framework to conduct rigorous feasibility analyses. Drawing on semantic search over curated scientific databases, online literature retrieval, and iterative evaluation, the system produces comprehensive feasibility reports detailing scientific and mathematical basis, suitability of employing Quantum/Quantum-Inspired computing methods, scientific evidence, limitations and risks, for the proposed use cases. By combining conversational problem scoping, algorithmic mapping, and autonomous scientific investigation, our platform enables organizations to make evidence-based, strategic decisions about quantum adoption.
The problem, and what does our new tool address?
Despite the fact that developing end-to-end, production-ready quantum solutions for business problems is not yet ready, recent research advances, such as progress on error-correction techniques and improved qubit stability, show clear movement toward practical deployment [1]. Because potential applications span numerous industries, the space for application software is far from saturated [2], offering significant opportunity but also considerable uncertainty. As a result, organizations often struggle to assess whether their computational challenges are well-suited to quantum approaches and how best to evaluate and validate such opportunities. Survey studies, such as [3], highlight that the complexity of quantum systems and software is among the primary barriers to broader adoption. This uncertainty slows progress and leaves decision-makers without the evidence needed to confidently justify investment.
Our tool, the QCentroid Multi-Agent AI Use Case Generator tool, was designed precisely to try bridging this gap. It transforms informal business conversations into structured, scientifically grounded quantum computing use cases. By combining conversational scoping, algorithmic mapping, and autonomous deep research, the system helps organizations cut through complexity and make data-driven decisions about quantum feasibility.
Technical Concepts
Before proceeding to explain each stage of the architecture, let´s briefly explain some technical concepts that are used in this architecture and will be mentioned along the article.
Prompting techniques: We applied several prompting techniques across different sections of the system. These included:
- Persona Prompting: Crafting prompts that assign the model a specific role or perspective.
- Chain-of-Thought (CoT) Prompting: This technique enables complex reasoning by guiding the model through intermediate reasoning steps [4].
- Few-Shot Prompting: In addition to task descriptions, the model is shown a few illustrative examples. This helps generalize to new tasks by following the demonstrated patterns [5].
Graphs: The architecture was implemented using LangGraph. At its core, LangGraph models workflows as graphs, where the behavior of agents is defined by three main components:
- States: Shared data structures that represent the current state of the application.
- Nodes: Functions that encode the logic of the agents.
- Edges: Functions that determine the next step to execute based on the current state.
Subgraphs: A subgraph is a graph encapsulated as a single node within another graph. In our architecture, the three main stages—Interview, Generator, and Deep Research—were encapsulated as subgraphs within the multi-agent system. This design offers several benefits:
- Modularity: Each subgraph can be developed and tested independently, having their own states. For example, we could test different types of users and business contexts in the interview subgraph, without needing to run subsequent subgraphs. The same applied to the other stages.
- Specialization: Agents can focus on specific domains. This opens opportunities to expand beyond quantum computing—for instance, by adding expert agents specialized in Machine Learning (ML).
- Control: Flow management is more straightforward. For example, in the Deep Research subgraph, we were able to execute parallel procedures efficiently thanks to the subgraph-based design.
Figure 1. Example of shared and private states in two subgraphs.
One example of the benefits of modularity is illustrated in Figure 1. The diagram shows how two subgraphs share the common “messages” state, while still maintaining their own private states: “generate_use_cases” for the Generator subgraph and “pinecone_search_query” for the Deep Research subgraph.
Context engineering: is about designing dynamic systems that provide the right information and tools in the right format, enabling Language Large models (LLM) to accomplish tasks effectively. Complex agents such as the architecture shown in this article, gather context from multiple sources— users, past interactions, tool calls, or external data—and must integrate these dynamically. Since LLMs cannot infer missing details, success depends on supplying accurate context and relevant tools. Equally important is format: clear, structured inputs enable the model to reason and act far more reliably than poorly organized or ambiguous data [6]. Context engineering was widely used in the research agent process, which we will discuss later.
Structured output: Refers to guiding the model to produce responses in a predefined format, it is important for context engineering to provide clear and structured information. To achieve this, a schema must first be defined—commonly using a JSON-like structure or libraries such as Pydantic. LangChain simplifies this process with the with_structured_output() method, which automatically binds the schema to the model and ensures the output is parsed correctly into the desired structure.
Map-Reduce: Advanced design pattern that allows parallel execution in <LangGraph, using the Send object.
Retrieval Augmented Generation (RAG): is the process of optimizing the output of a large language model, so it references an authoritative knowledge base outside of its training data sources before generating a response [7].
Note: In this article, the words subgraphs and agents can be used interchangeably, since in our architecture each subgraph acts as an agent that can decide the control flow of the application.
The Architecture
Figure 2. AI Multi-Agent Quantum computing Use Case Generator architecture
This section provides a comprehensive overview of our Multi-Agent Quantum Computing Use Case Generator tool. The architecture is divided into three-stages: Interview, Generator, and Deep Research, as seen in Figure 2.
In this architecture, each stage corresponds to a subgraph that acts as an independent agent and has its own states. The generator and the Deep Research agents, receive the conversation summary and quantum computing use cases from previous stages, respectively. However it is worth noting that these subgraphs also share some other states, as for example the states of the conversation messages between the user and the interview agent.
Stage 1: Interview – Scope
The foundation of any relevant use case is a deep understanding of the problem. Our initial stage is an interactive interview agent designed to map the user’s specific context.
Figure 3. Example of user interaction.
The agent in this stage contained three main nodes: get_human_feedback, conduct_interview, evaluate_interview. These nodes were connected through edges in a loop, to perform the following tasks:
- User Data Collection: In this node the agent engages the user to collect critical data points: strategic objectives, operational challenges, existing computational bottlenecks, and unexplored opportunities. We used the CoT Prompting technique, to guide the agent in the collection of the user information. As mentioned before, this process is an iterative loop, continuing until the information collected is established for analysis. Figure 3 illustrates an example conversation in which can be appreciated how the agent scopes the user’s business.
- Messages evaluation: The purpose of the agent in this node is to actively guide the conversation, ensuring alignment with topics amenable to quantum computing or quantum-inspired solutions. It employs conditional logic to steer the dialogue towards quantifiable problems, such as optimization, simulation, or machine learning challenges, and is programmed to conclude interviews that diverge from these domains.
- Brief Generation: The stage culminates in a structured “Conversation Summary,” a formalized document that serves as the input for the subsequent stage.
Stage 2: The Generator – Generation of Quantum Computing Use Cases
With a clearly defined scope, the Generator agent synthesizes this information to propose initial solutions.
- Interview Evaluation: In this node the agent analyses the Conversation Summary input to confirm the feasibility of applying quantum computational methods. This acts as a first validation gate before generating the quantum use cases.
- Quantum Computing Use Cases Generation: A “Quantum Expert” LLM then synthesizes the user’s business information with its knowledge base of quantum computing algorithms. It generates a preliminary list of use cases, mapping business problems like logistics optimization to algorithms such as the Quantum Approximate Optimization Algorithm (QAOA) or Variational Quantum Eigensolver (VQE), or material science simulations to quantum chemistry algorithms.
Stage 3: Deep Research – Autonomous, Multi-Agent Feasibility Analysis
This final stage is the centerpiece of our system: a parallelized AI multi-agent architecture, inspired by the open Deep Research architecture [8], which conducts a rigorous academic and technical investigation for each proposed use case. Its purpose is to assess the feasibility of applying quantum computing and quantum-inspired algorithms to the user’s business. This serves as a second validation gate before producing scientifically grounded quantum computing use cases.
Figure 4. Parallel Deep Research
The parallel procedure is illustrated in Figure 4. We apply the map-reduce design pattern to branch the workflow, enabling parallel execution for each use case. The Generator agent forwards the list of use cases to the next subgraph using LangGraph’s Send object. For each use case, a specialized agent is instantiated, allowing research tasks to be carried out independently, and more efficiently, at the same time.
We apply context engineering [9] in different procedures to extract key information, keep relevant content and sources, and compress research results. We followed established guidelines for model families such as GPT-4.1[10], ensuring more effective use of the model’s capabilities. During this procedure we also applied structured outputs to maintain clear and structured data in the flow execution of the nodes.
As seen in the third block of Figure 1, the process begins with the Deep Research Loop:
- The Interviewer: This new LLM interviewer acts as a Socratic inquisitor. Based on the proposed use case (e.g., “Portfolio Optimization using Quantum Annealing”) and the user’s context (e.g., a financial services firm concerned with volatility), it formulates precise research questions. For example: “What is the latest research on the performance of D-Wave’s annealers for QUBO problems with constraints similar to our user’s portfolio?”
- The Quantum Computing Expert: This agent takes the questions and queries both internal and external knowledge sources. To ensure the LLMs correctly understood their roles during the interview, we applied persona prompting. Specifically, we defined two personas: an Interviewer persona, as described in the previous section, and a Quantum Expert persona, which responds to all questions using the information retrieved by the tools.
- Database and Tools: The query is forwarded to a specialized vector database populated with embeddings derived from hundreds of recent arXiv preprints, peer-reviewed journal articles, and quantum algorithm literature. To enable this, we employ a RAG pipeline, leveraging dense vector indexes in Pinecone that support high-dimensional semantic similarity search. In parallel, the agent integrates external tools such as Tavily to perform live web searches, ensuring access to the most up-to-date scientific studies and complementing the static knowledge base with real-time information.
- Source Evaluation: An evaluation layer filters the retrieved sources by the quantum computing expert, ranking them based on relevance to the Interviewer query, and applicability to the user’s specific industry. Irrelevant or sources are discarded.
This loop iterates, refining the search and building a rich corpus of relevant knowledge.
Figure 5. Example of generated use cases and feasibility reports.
- Deep Research Evaluation. At the end of the loop, the AI multi-agent system conducts a final review, examining the collected evidence and determining the feasibility of applying the quantum use cases to the user’s business. The review applies constraint checks (e.g., data availability, problem size, hardware/runtime requirements, regulatory limits), compares against baselines, and outputs a confidence-weighted verdict. Few-shot prompting is employed to guide the evaluator by providing contrasting examples of feasible and non-feasible cases, ensuring more consistent and accurate judgments.
- Report Writing: The process culminates in a comprehensive feasibility report (see an example in Figure 5). This is not a simple summary but a structured scientific document containing:
- Scientific and Mathematical Basis: A brief mathematical description of the quantum computing algorithms used for the use case proposed.
- Limitations and Risks: A clear-eyed view of current hardware limitations, algorithmic noise, and implementation challenges.
- Conclusion: A final verdict on the feasibility and potential benefits for the user’s business.
- Scientific Evidence and Literature Review: A curated list of supporting studies with citations.
Our tool is designed to be more than a simple tool generator; it is an automated quantum computing research assistant that empowers organizations to make informed, data-driven decisions about their entry into the quantum computing landscape.
What´s next?
We are still cooking …
At QCentroid, we have the goal of accelerating the large-scale adoption of quantum computing. We’ve got several planned and in-progress features to achieve this end-to-end workflow: from defining a use case to developing it into test-ready code. You’ll start with a simple conversation and end with a working experimental prototype, supported by tools such as a mathematical problem-definition generator, a data generator, and pseudocode-to-code generation—though that’s just the beginning; the rest you’ll have to wait and see.
