Research Focus

Agentic AI

Autonomous systems based on LLMs for complex environments

Explore content Back to GPAIS

What is Agentic AI?

Agentic AI is the paradigm in which large language models (LLMs) stop acting as simple text generators to become autonomous agents capable of planning, reasoning, using tools, and executing complex multi-step tasks.

The term Agentic is the adjective that encompasses all agents with varying degrees of autonomy, from assistants requesting confirmation at each step to fully autonomous systems. This spectrum of autonomy makes Agentic AI a particularly powerful strategy in complex environments.

Agentic vs. Non-Agentic Workflows

The fundamental difference lies in how the interaction with the model is organized.

➡️Non-Agentic Workflow

The model receives a query and produces a response without the ability to review, reflect, or interact with external tools on its own. Suitable for simple and well-defined tasks.

🔄Agentic Workflow

The model operates in iterative loops: it plans, acts, observes results, and revises its plan. It can decompose into specialized sub-agents and delegate tasks, resembling the work of an expert human team.

Degrees of Autonomy

Agentic systems are not all-or-nothing: there are multiple levels depending on how much control the human operator retains.

Low

Human-in-the-Loop

The human operator approves each significant action. Maximum safety but lower efficiency. Appropriate for high-criticality environments where an error can be irreversible.

Medium

Selective Oversight

The agent acts autonomously in routine tasks and escalates to human review only for low-confidence or high-impact situations. Balance between efficiency and safety.

High

Full Autonomy

The system manages complete cycles without human intervention. Requires solid guarantees of safety, traceability, and robustness. Still a long-term objective.

Why Do LLMs Fail?

Understanding the limits of LLMs is the starting point for building reliable agentic systems in complex environments.

Lack of Reasoning

LLMs learn statistical correlations, not cause-effect relationships. Following the causal ladder, they operate primarily at an association level, without innate ability to reason about interventions or counterfactuals.

Hallucinations and Factuality

Models can generate plausible but incorrect responses with high confidence. In critical facilities, a fabricated response can lead to dangerous operational decisions, making factuality a non-negotiable requirement.

Static Knowledge

LLMs have a training cutoff date and don't access real-time data. In complex and critical facilities, operational parameters and technical documentation change continuously, requiring dynamic contextualization strategies.

Highly Specialized Domain

The technical vocabulary, conceptual relationships, and safety protocols of a critical facility are scarce in the pretraining data of general LLMs, degrading their performance.

Our Research

In the GPAIS group at IAFER chair, we work at the intersection of Agentic AI and the real needs of critical scientific facilities. The synergy between artificial intelligence and highly complex scientific environments represents a unique opportunity to advance both fields.

🚀 Active Research Lines

Tool-augmented reasoning: Integration of symbolic reasoning systems with LLMs for verifiable decision-making in critical facilities
Grounding and retrieval-augmented generation (RAG): Methods to ground LLM outputs in domain-specific documentation, ensuring factuality and traceability
Agent planning and multi-step reasoning: Decomposition of complex tasks into subtasks with explicit verification and rollback mechanisms
Safety and auditing: Mechanisms to ensure agent decisions are transparent, auditable, and compliant with operational safety protocols

📚 Publications

Loading publications...

Applications in Critical Facilities

Agentic AI has a significant impact in environments where technical complexity and safety requirements make purely manual operation infeasible. Our research has direct impact on:

Operator Assistance: Natural language queries of manuals, protocols, and historical facility records
Knowledge Management: Efficient indexing and retrieval of constantly evolving technical documentation
Decision Support: Expert systems that synthesize information from multiple sources to recommend actions to operators
Procedure Verification: Agents that check consistency and factuality of standard operating procedures