Agents that learn: a soft actor-critic approach

About this session

At AI Week 2026, we will present an architecture for autonomous, adaptable agentic systems oriented toward continuous learning. At the core of the solution, reasoning is separated from execution: a first multi-agent system composed of “Actors” and “Critics” collaborates in an adversarial setup to define the action strategy, also leveraging the history of past executions. A second multi-agent system, organized hierarchically, is responsible for execution and real-world implementation through specialized agents and operational functions, known as “Tools”. The system implements a reinforcement learning paradigm, where evaluating agents compare the generated outcome with user expectations, feeding a global and personalized memory that ensures continuous performance improvement and prevents the repetition of errors. The infrastructure, which is portable on-premise, guarantees full data sovereignty and security, featuring a modular and LLM-agnostic design that eliminates vendor lock-in risk. Finally, the multi-agent architecture ensures maximum explainability and transparency of decision-making processes, enabling the solution to evolve and scale in close collaboration with end users through co-design.

Speakers

LG

Lorenzo Gagliardi

Lead Data Scientist TIM Enterprise, TIM Enterprise


TM

Tommaso Colombo Maria

Tech Area Manager TIM Enterprise, TIM Enterprise