Beyond the Hype: Rebuilding Agent AI with Data at its Core

Author: Denis Avetisyan

Current agentic AI systems struggle with real-world enterprise applications, and this review proposes a fundamental shift in architecture to prioritize structured data access and reliable execution.

The RUBICON architecture facilitates query processing through an AQL-based system, establishing a foundational framework for data retrieval and manipulation.

This paper details RUBICON, a data-centric architecture leveraging AQL for multi-source reasoning and knowledge integration in agentic AI systems.

Despite recent advances in artificial intelligence, enterprise applications often struggle not from a lack of reasoning ability, but from fragmented and inaccessible data. This challenge is addressed in ‘An Alternate Agentic AI Architecture (It’s About the Data)’, which argues that current agentic AI systems, heavily reliant on large language models, misdiagnose the core problem-data integration-by treating enterprise systems as unstructured text corpora. The paper introduces RUBICON, a data-centric architecture utilizing an Agentic Query Language (AQL) to enforce schema alignment, access control, and deterministic execution across heterogeneous data sources. By prioritizing explicit query structure over opaque LLM orchestration, can we finally unlock the promise of trustworthy and scalable AI within complex enterprise environments?

The Limits of Pattern Matching

Large Language Models demonstrate remarkable proficiency in identifying and replicating patterns within data, a capability stemming from their training on massive datasets. However, this reliance on scale-simply increasing the size of the model and the data it processes-introduces limitations when confronted with tasks demanding genuine complex reasoning or reliable access to specific information. The models often struggle to extrapolate beyond the patterns they’ve memorized, exhibiting brittle performance when faced with novel situations or nuanced queries. Furthermore, retrieving accurate data isn’t inherent to the architecture; LLMs can confidently generate plausible-sounding but factually incorrect responses-a phenomenon known as hallucination-highlighting the need for mechanisms that ground their reasoning in verifiable sources and ensure data integrity, rather than solely relying on statistical correlations.

Conventional techniques like Text-to-SQL, designed to translate natural language questions into database queries, frequently falter when deployed within the complexities of real-world enterprise environments. These systems struggle with nuanced language, ambiguous phrasing, and the evolving schemas common in large databases, leading to inaccurate results and unreliable data retrieval. This brittleness significantly impedes the creation of truly functional Agentic AI – systems intended to autonomously access information and perform tasks – because their ability to reliably ground reasoning in accurate data is fundamentally compromised. Consequently, a reliance on these traditional methods introduces a critical bottleneck, preventing Agentic AI from achieving the robustness and dependability required for practical application and scalable deployment.

Current architectures heavily reliant on Large Language Models (LLMs) face significant hurdles due to the computational expense and delays inherent in processing extensive token sequences. Performance benchmarks reveal a critical limitation: LLM-centric systems consistently achieve zero percent accuracy when tasked with complex, multi-source data queries. This starkly contrasts with data-centric approaches, which, by prioritizing efficient data retrieval and structured information access, demonstrate the capacity to attain one hundred percent accuracy on the same benchmarks. These findings suggest that simply scaling LLMs is not a viable path toward robust Agentic AI; instead, a fundamental shift toward architectures that minimize token processing and maximize data efficiency is crucial for unlocking truly functional and reliable artificial intelligence systems.

A Foundation of Structured Knowledge

A Data-Centric Architecture represents a shift in agentic AI development, moving away from a primary focus on Large Language Model (LLM) reasoning as the sole driver of intelligence. This approach prioritizes the organization, accessibility, and utilization of structured data as the foundation for AI systems. Rather than relying on LLMs to extract knowledge from unstructured text during each interaction, a Data-Centric Architecture emphasizes pre-processing and integrating data from various sources – including enterprise data warehouses and publicly available databases – into a readily accessible format. This minimizes the need for extensive LLM token usage during inference, directly addressing limitations in cost, latency, and scalability inherent in LLM-centric approaches that depend heavily on in-context learning and complex prompting strategies.

A data-centric architecture for agentic AI necessitates the integration of structured data from a variety of sources. This includes internal Enterprise Data Warehouses, which often contain proprietary and curated information critical for specific business functions, and external, publicly available resources such as Wikipedia, providing broad general knowledge. Connecting these disparate data sources requires robust data pipelines and standardized data formats to ensure consistency and accessibility. Successful integration allows the AI agent to leverage a significantly larger and more diverse knowledge base than is typically embedded within the LLM itself, enabling more accurate and contextually relevant responses. The architecture supports data retrieval based on specific agent requests, delivering only the necessary information for task completion.

A key benefit of data-centric architectures is the significant reduction in Large Language Model (LLM) token usage. Traditional LLM-centric approaches, such as ReAct prompting, often require substantial input tokens – tests with GPT-5-mini have demonstrated consumption ranging from 20,000 to 46,000 tokens per interaction. Conversely, systems leveraging pre-processed, relevant data, like RUBICON, achieve comparable results with drastically fewer tokens, typically under 80. This reduction in token usage directly translates to lower operational costs and decreased latency, enabling faster response times and more scalable deployments of agentic AI systems.

RUBICON: Bridging Data and Reasoning

RUBICON’s Data-Centric Architecture is facilitated by the Agentic Query Language (AQL), a system designed for efficient data retrieval and integration. AQL operates by translating natural language queries into structured data requests, allowing RUBICON to directly access and combine information from multiple sources without relying on iterative prompting or large language model interpretation. This direct access minimizes latency and reduces the potential for errors inherent in LLM-based approaches, as AQL focuses on precise data manipulation rather than probabilistic reasoning. The language is designed to be both human-readable and machine-executable, enabling developers to easily define and manage data flows within the system.

RUBICON utilizes wrappers to establish a uniform interface for interacting with heterogeneous APIs. These wrappers abstract the specific implementation details of each API, providing a standardized method for data retrieval and manipulation. This standardization allows RUBICON to connect to and integrate data from a variety of sources – including databases, web services, and external tools – without requiring modifications to the core system. The wrappers handle authentication, request formatting, and response parsing, ensuring consistent data access regardless of the underlying API’s protocol or structure. This approach simplifies the integration process and enhances the system’s adaptability to new data sources.

The RUBICON Query Optimizer functions by analyzing incoming queries and generating multiple potential execution plans, which are then evaluated based on estimated cost and resource utilization. This process involves considering factors such as data source characteristics, available indexes, and the complexity of individual operations. The optimizer selects the plan with the lowest projected cost, minimizing latency and overall resource consumption during query execution. By dynamically adapting to query structure and data distribution, the optimizer ensures efficient data retrieval and processing, contributing to RUBICON’s performance gains over LLM-centric approaches.

Performance benchmarks demonstrate that RUBICON consistently achieves 100% accuracy on multi-source queries, a result not currently replicable by Large Language Model (LLM)-centric approaches. Comparative analysis indicates a substantial reduction in computational load; RUBICON requires an average of 2.0 tool calls to complete a query, whereas Gemini-3-flash-preview averages 22.71 tool calls for the same tasks. This difference in tool call frequency directly correlates with improved efficiency and reduced latency, highlighting the benefits of RUBICON’s Data-Centric Architecture and Agentic Query Language (AQL) in complex data retrieval scenarios.

Empowering Intelligence Through Structure

RUBICON significantly enhances Retrieval-Augmented Generation (RAG) by shifting the focus from raw text to meticulously organized knowledge. Rather than tasking Large Language Models (LLMs) with deciphering unstructured data, RUBICON delivers information already parsed, categorized, and linked – essentially providing LLMs with a pre-digested understanding of the relevant facts. This pre-processing dramatically improves the efficiency and accuracy of RAG systems, allowing them to swiftly identify and utilize the most pertinent information for any given query. The result is a substantial reduction in ‘noise’ and ambiguity, empowering LLMs to generate more coherent, reliable, and insightful responses – a crucial step towards building truly knowledgeable and capable artificial intelligence.

The RUBICON architecture distinguishes itself through a deliberate emphasis on computational efficiency, actively minimizing the need for ever-larger language models. Rather than relying on brute-force scaling – simply increasing parameters and data volume – the system prioritizes data quality and pre-processing to extract meaningful structure. This approach yields significant reductions in computational cost and, consequently, environmental impact; complex tasks can be accomplished with considerably less energy expenditure. By intelligently structuring information before it reaches the language model, RUBICON demonstrates that impactful AI doesn’t necessarily require exponentially growing resources, offering a path towards sustainable and accessible artificial intelligence.

The architecture demonstrates a pathway towards genuinely intelligent agents, moving beyond simple pattern recognition to enable complex reasoning and robust decision-making capabilities across a wide spectrum of applications. Unlike conventional large language model (LLM) approaches, which often struggle with tasks requiring structured understanding, this method achieved a perfect 100% accuracy rate on benchmark tests specifically designed to assess such skills-tests where LLM-centric systems consistently failed. This performance suggests that by prioritizing data quality and structured knowledge integration, systems can overcome limitations inherent in solely relying on the scale of LLMs, opening possibilities for reliable AI in areas demanding precision and logical inference.

The architecture detailed within prioritizes structured data access as the foundation for reliable agentic behavior, a principle echoing Blaise Pascal’s observation: “The eloquence of a man does not depend on the number of his words, but on the force and precision of his thought.” Just as impactful communication requires focused expression, robust AI agents necessitate precise data handling. The RUBICON framework, by emphasizing deterministic execution through AQL and multi-source reasoning, seeks to move beyond the probabilistic nature of LLM-centric approaches. This focus isn’t merely about technical efficiency; it’s about building systems where behavior predictably stems from clearly defined inputs and a well-understood structure – a holistic approach to agent design.

Where Do We Go From Here?

The pursuit of intelligence, artificial or otherwise, consistently reveals a humbling truth: the elegance of a system resides not in the sophistication of its central component, but in the integrity of its foundations. The emphasis on increasingly large language models, while producing superficially impressive results, skirts the fundamental issue of reliable knowledge integration. The architecture presented here, RUBICON, attempts to address this by prioritizing structured data and deterministic execution, but it does not, and cannot, resolve the inherent messiness of real-world information. The next logical progression necessitates a deeper exploration of data governance-not merely access, but provenance, validation, and the explicit representation of uncertainty.

AQL, as a query language for agentic systems, is a necessary step, but it’s only as robust as the data it interrogates. Future work must consider how to bridge the gap between the continuous, probabilistic world and the discrete, logical demands of computation. This is not a technical problem alone; it’s a question of epistemology-how do we represent what we know, and, crucially, what we do not know?

Ultimately, the challenge lies in accepting that true intelligence isn’t about mimicking human cognition, but about building systems that are demonstrably reliable. The field must resist the allure of complexity for its own sake, and instead focus on minimizing dependencies and maximizing transparency. A system built on clear boundaries, even if limited in scope, will always outperform one drowning in opaque, unbounded possibilities.

Original article: https://arxiv.org/pdf/2604.21413.pdf

Contact the author: https://www.linkedin.com/in/avetisyan/

The Limits of Pattern Matching

A Foundation of Structured Knowledge

RUBICON: Bridging Data and Reasoning

Empowering Intelligence Through Structure

Where Do We Go From Here?

See also: