TLDR: Agentics is a new framework from IBM Research that enables Large Language Models (LLMs) to work effectively with structured data. It introduces “logical transduction,” a data-centric process where stateless agents transform data between defined types based on schemas. This approach improves accuracy, scalability, and robustness in tasks like domain-specific question answering, text-to-SQL, and prompt optimization, by focusing on data modeling rather than complex prompt engineering. The framework also leverages an asynchronous Map/Reduce model for efficient, parallel execution.
Large Language Models (LLMs) have transformed how we interact with AI, excelling at understanding and generating natural language. However, their strength in unstructured settings often falters when dealing with structured data—information defined by explicit schemas, identifiers, and deterministic rules, common in databases and enterprise systems. Traditional agentic AI frameworks often try to bridge this gap by embedding structured data into natural language prompts, a method that can be fragile and inefficient for tasks requiring high precision, reproducibility, and scale.
IBM Research has introduced a novel framework called Agentics, designed to address these limitations. This modular system aims to build agent-based AI that can perform structured reasoning and generalize across complex data types. Agentics shifts the focus from crafting elaborate prompts to modeling data itself, offering a declarative language where LLMs define and compose data types through a process called logical transduction.
Understanding Logical Transduction
At the heart of Agentics is the concept of logical transduction. Imagine data as a raw material that needs to be reshaped. Logical transduction is the process of transforming a data object of one type (Type X) into another (Type Y) by intelligently inferring and populating fields in Type Y based on the information in Type X, all while adhering to the constraints of a target schema. This process is executed by lightweight, stateless agents that are designed to be composable, parallelizable, and verifiable.
Unlike conventional agentic AI, which often relies on anthropomorphic metaphors like memory, tools, and conversation, Agentics views agents as simple, stateless transducers. This means each transformation is independent of previous or future operations, enabling highly asynchronous and parallel execution. Subject Matter Experts (SMEs) can define workflows by specifying data schemas (using Pydantic) and linking them via these transductions, creating a no-code approach to data ingestion and transformation.
Scalable Workflows with Asynchronous Map/Reduce
To handle enterprise-scale data, Agentics incorporates an asynchronous Map/Reduce programming model. This extends the traditional MapReduce paradigm by allowing mapping and reduction operations over Agentic structures to run asynchronously. This capability is crucial for blending LLM inference with program logic in a scalable manner, ensuring efficient processing of large datasets.
The framework’s implementation centers around an `AG` meta-class, which binds a Pydantic type with its associated data states, and a `PydanticTransducer`. This transducer is a stateless agent that generates valid instances of the target data type from textual input, ensuring structured and reliable outputs. The framework uses a simple `<<` operator for performing these transductions, making complex data transformations intuitive to express.
Empirical Evidence and Performance
Agentics has been rigorously tested across various challenging tasks, demonstrating significant improvements:
- Domain-Specific Multiple-Choice Question Answering (MCQA): On the FailureSensorIQ benchmark, Agentics substantially improved accuracy across all evaluated models, with smaller models like Qwen3-8B seeing a remarkable +14.32% gain. It also showed robustness against question perturbations and achieved significant speedups through parallel batch execution.
- Automatic Prompt Optimization: The framework naturally supports prompt optimization techniques, leading to improved test scores on datasets like GSM8K and FailureSensorIQ. Its asynchronous execution capabilities also resulted in substantial runtime gains during the optimization process.
- Text-to-SQL: By composing multiple transduction steps, such as incorporating few-shot examples, schema linking, and optimized prompts, Agentics achieved up to a 10.33% increase in execution match results on the Bird-bench dataset, showcasing its ability to handle complex semantic parsing tasks.
Also Read:
- Agent Foundation Models: A Unified Approach to AI Problem Solving
- Intelligent Configuration Optimization with Language-Guided Tuning
A New Paradigm for AI Frameworks
Agentics stands apart from other popular agentic AI frameworks like LangGraph, CrewAI, AutoGen, and DSPy. While these systems often focus on agent behavior and control flow, Agentics prioritizes semantic modeling of data and logical transduction. It offers stronger guarantees through type validation, reduces reliance on extensive prompt engineering, and treats LLM calls as native asynchronous functions, simplifying development while maintaining flexibility and composability.
By redefining how agents interact with data through a declarative, type-driven approach, Agentics offers a powerful new perspective on building intelligent systems that can effectively handle structured data workflows at scale. For more details, you can read the full research paper here.
