How can we teach an LLM agent to improve over time without the massive cost of retraining it? In this talk, I present Memento, a framework that shifts the burden of learning from neural parameters to episodic memory. We argue that for an agent to be truly generalist, it must learn like a human: by recalling analogous situations from the past. Memento implements this via Case-Based Reasoning (CBR), allowing a "Planner-Executor" architecture to search a growing library of its own experiences to solve complex, long-horizon tasks. I will walk through the technical architecture—specifically our Memory-augmented MDP—and demonstrate how Memento achieves state-of-the-art results on benchmarks like GAIA and DeepResearcher. 

Traditional reinforcement learning produces super-human reactive systems, capable of mastering games through trial-and-error with scalar reward. Yet, a critical dimension of intelligence remains unexplored: the ability to form a deep, deliberate, causal understanding of the game itself. This talk outlines a new direction for training AI with the language of action—agents that not only play a game proficiently but also genuinely understand it. The journey begins with ChessGPT, a model that learns deliberative thinking from human collective experience. By co-training on millions of chess games and their associated strategic analyses, this model learns to connect actions to their underlying rationale, and forms a rich, structural understanding can be distilled from existing knowledge bases. The critical next step towards autonomy is then presented as Natural Language Reinforcement Learning (NLRL). By taking analogy from RL, NLRL presents how agent's can learn deliberative thinking through the agent's own life-stream of experience. 

Multi-agent systems are increasingly deployed in dynamic environments where agents must coordinate their strategies to achieve optimal outcomes. However, conventional game-theoretic approaches often struggle with issues such as multiple Nash equilibria, limited agent controllability, and scalability constraints. In this talk, I will introduce a causal intervention perspective to address these challenges. By leveraging structural causal games and multi-agent influence diagrams, we reformulate coordination problems as inference tasks on causal graphs. This enables localized interventions - so-called pre-policy interventions, that steer agent behavior toward globally desirable equilibria, even when full control over all agents is infeasible. I will discuss how these methods align with human preferences, break cyclic dependencies in strategy updates, and enhance robustness in both simulated environments and real-world applications such as human-AI collaboration. This causal lens offers a principled framework for influencing emergent outcomes in complex multi-agent systems.

This talk will introduce a new foundation for AGI agents grounded in human experiential learning, drawing directly from Kolb’s theory and enriched by Vygotsky’s notion of scaffolded development. I present Agent K, an autonomous system that learns by iteratively experiencing, reflecting, conceptualising, and experimenting — not only in controlled environments, but also across open-ended tasks. Applied to end-to-end data science, Agent K achieves competitive performance on a diverse suite of Kaggle competitions, reaching 9 gold, 8 silver, and 12 bronze medal–level results, including 4 gold and 4 silver in prize-awarding challenges. Beyond data science, I demonstrate how such frameworks extend to embodied AI, enabling agents that adapt across robotic settings and modalities. Taken together, these results highlight a path toward generalist AI systems that learn and develop more like humans — through structured experience, guided exploration, and continual self-improvement.

This talk introduces the CAMEL framework through a focused overview of agent evolution, highlighting how modern LLM-driven agents advance task automation, world simulation, and data generation. It outlines the motivations behind CAMEL’s design and how multi-agent workforce address emerging challenges in building reliable, scalable, and coordinated Agentic future. We will further use three representative use cases to illustrate how our product, Eigent, brings multi-agent systems into real-world production settings and how our open-source ecosystem connects academic research with industrial adoption.

We will open the floor to have a discussion about fundamental questions, future directions, and industrial applications of LLM-based multi-agent systems.