AI Agents: Past, Present, and Future 

AI agents are no longer a futuristic concept. They’re here, and they’re reshaping science, medicine, and society. But what makes an agent truly intelligent? Beyond a simple chatbot, a new generation of agents can model reality, generate complex code, formalise knowledge, and collaborate with experts. In our lab, we’ve been building these kinds of agents for the past three years, laying the foundations for what is fast becoming the next frontier of AI. 

This trajectory began with our Reality-Centric AI vision, written before the rise of large language models and before agents entered the spotlight. The article posed a challenge: could AI move beyond benchmarks and toy problems to engage with the full messiness of the real world? The response has exceeded even our wildest hopes. Today, agents are emerging as the clearest embodiment of that reality-centric vision. 

In this post, we’ll reflect on that journey: how our lab anticipated and shaped the rise of agents, the diverse systems we’re building right now across seven pillars of research, and what comes next: networks of agents working together to accelerate discovery and redefine what it means for AI to be a true collaborator. 

Before we dive in, let’s clarify what we mean by an AI agent. 
An agent is an AI system that can perceive information, reason about it, and take actions to achieve a goal in a dynamic environment. Unlike a static tool, an agent can adapt, interact, and collaborate – becoming an active participant in problem-solving rather than a passive responder. 

In our lab, we extend this idea further: agents that don’t just assist, but actively collaborate in scientific discovery and medical decision-making, embodying the reality-centric vision that has guided our work. 

The Past: Pioneering the Foundations of AI Agents 

The rise of AI agents may feel sudden to many, but it reflects a vision that has been years in the making, and decades in the shaping.  

Building on more than 20 years of experience in building reality-centric AI in healthcare, education, biology, communication networks, smart cities, smart grids, finance, crowdsourcing, labour markets, our lab set out three years ago with the Reality-Centric AI vision a call for systems that could move beyond passive assistance to become active collaborators, capable of grappling with the complexity and unpredictability of the real world. 

At the time, this stance was far from mainstream. Much of AI research was centred on benchmarks, and incremental progress. The framework challenged a different path: that the true test of AI would be its ability to deal with messy, dynamic reality. That idea guided our first explorations into agents – systems that could generate code, discover scientific models, or build digital twins – early examples of what reality-centric intelligence might look like in practice.  

Looking back, what once seemed controversial has now become a shared priority. The conversation in AI has shifted dramatically, and thousands of researchers worldwide are now working on agents. Agents have become the clearest embodiment of the reality-centric vision, in ways that have grown beyond what we could have anticipated at the time. 

While the broader AI community has only recently focussed on agents, our lab has been at the forefront of this area for more than three years. Long before today’s industry spotlight, we were exploring a fundamental question: What if AI systems could evolve from being passive tools to active collaborators? 

This question became the starting point for one of our earliest and most distinctive ideas: the Genies framework. Conceived nearly 3 years ago, Genies outlined a vision for AI agents that went far beyond what others were imagining at the time. While most of the field was still focused on narrow, task-specific assistants, we proposed a new generation of agents that could combine multiple capabilities, adapt across tasks, and act as trusted partners in research. Genies embody abilities that most agents today still lack – including creativity, continuous learning – making them a comprehensive blueprint for the future of AI agents. These ideas provided the intellectual groundwork for a line of inquiry that would soon define a major direction in AI. 

What began as a bold vision soon turned into concrete research. We started building prototypes that tested these ideas in practice – agents that could code, model reality, and uncover scientific systems. Each project was a way of asking: what does it really mean for AI to collaborate with us, rather than simply respond to us? 

From there, we began publishing some of the first academic demonstrations of agents at top international conferences. Each project established a different dimension of what agents could do: 

• L2MAC: Large Language Model Automatic Computer for Extensive Code Generation (ICLR 2024) [https://arxiv.org/abs/2310.02003] 
  We introduced L2MAC as a general-purpose stored-program automatic computer, with an instruction registry and file store that enabled agents to autonomously generate and refine code at scale. 
• Automatically Learning Hybrid Digital Twins of Dynamical Systems (NeurIPS 2024) [https://arxiv.org/abs/2410.23691] 
  This work established the concept of agents as “digital twin builders,” capable of dynamically simulating individuals to explore interventions and predict outcomes. 
• Data-Driven Discovery of Dynamical Systems in Pharmacology using Large Language Models (D³ framework) (NeurIPS 2024) [https://openreview.net/forum?id=KIrZmlTA92] 
  We demonstrated that agents could reason directly about scientific data, discovering complex dynamical systems in pharmacology and beyond without heavy human intervention. 
• Autoformulation of Mathematical Optimization Models Using LLMs (ICML 2025) [https://arxiv.org/abs/2411.01679] 
  We showed how agents could formalize reasoning processes, automatically translating human intuition into rigorous optimization frameworks. 
• Active Task Disambiguation with LLMs [https://arxiv.org/abs/2502.04485]  
  We developed agents that can actively disambiguate ambiguous instructions, clarifying tasks through interaction to ensure more accurate and reliable outcomes. 
• Matchmaker: Self-Improving Large Language Model Programs for Schema Matching (NeurIPS 2024) [https://arxiv.org/abs/2410.24105]  
  This work introduced Matchmaker, an agent framework that continuously improves its ability to align and reconcile data schemas, enhancing data integration at scale. 
• Context-Aware Testing: A New Paradigm for Model Testing with Large Language Models (NeurIPS 2024) [https://arxiv.org/abs/2410.24005]  
  We proposed a context-aware testing paradigm, where agents use LLMs to generate adaptive, scenario-driven tests that better evaluate model robustness. 
• Continuously Updating Digital Twins using Large Language Models (ICML 2025) [https://arxiv.org/abs/2506.12091]  
  Extends our vision of how agents can update digital twins in real time. 
• G-Sim: Generative Simulations with LLMs and Gradient-Free Calibration (ICML 2025) [https://openreview.net/forum?id=PvkO6rIixC] 
  Advances the development of world models, showing how LLMs can generate and calibrate scientific simulations. 

Together, these projects marked some of the first demonstrations of what agents could achieve in practice. By translating the ideas of the Reality-Centric AI vision into concrete systems, they helped lay the foundations for a research direction that is now flourishing across the field. What began as an ambitious vision has become a shared agenda, with agents moving rapidly from concept to reality. 

The Present: From Vision to Reality 

What started with the Reality-Centric AI vision and our first agent prototypes has expanded far beyond our lab. 

Building on this foundation, our lab now leads seven key pillars of AI agent research. These pillars represent the domains where agents can have the most transformative impact on science and medicine, and where our group has made distinctive contributions. 

Each of these agent pillars also maps directly onto our Reality-Centric AI (RCAI) agenda, which sets out eight principles for building AI that can truly engage with the real world: 

• Pillar 1 – Identifying how to model the world: Our work on Autoformalism provides systematic ways to formalise and abstract complex real-world problems into rigorous optimisation frameworks. 

• Pillar 2 – Operating with real-world data: Our Data-centric AI projects (such as Matchmaker, SynthCity, and SynthCraft) tackle the challenges of noisy, biased, incomplete, or privacy-constrained data. 

• Pillar 3 – Determining what data to acquire: Research on Partially Observable Cost-Aware Active-Learning (POCA) (NeurIPS 2024) shows how agents can reason about what data is most valuable or necessary for a task, rather than assuming “more is better.” 

• Pillar 4 – Adapting post-deployment: Self-aware testing (NeurIPS 2024) demonstrates how agents can detect distribution shifts and adapt or at least flag when predictions are unreliable. 

• Pillar 6 – Respecting human constraints: Active Task Disambiguation (ICLR 2025) highlights how agents can resolve ambiguity in collaboration with humans, while research on human alignment ensures (ICLR 2025) that AI systems remain consistent with human goals, ethical principles, and regulatory constraints. 

Among these six pillars, Copilot agents deserve particular attention. Unlike traditional AI tools that replace human effort in narrow tasks, copilots are designed to amplify human expertise, helping researchers, clinicians, and decision-makers do things they could not do alone. 

Our lab has pioneered several copilots, including: 

• CliMB: An AI-enabled partner that guides clinician–scientists through the full data science pipeline, from exploration and engineering to model building and interpretation, using a no-code, conversational interface. 

• CliMB-DC: A data-centric extension of CliMB that allows experts to curate and refine datasets, ensuring models trained on complex clinical data are more reliable and trustworthy. 

• Causal Copilot: An assistant that supports reasoning about cause-and-effect in messy real-world data, helping researchers uncover confounders, test counterfactuals, and build causal models. 

• SynthCity/SynthCraft: Tools for generating high-quality synthetic data; SynthCity is a library focused on privacy, fairness, and augmentation, while SynthCraft provides an interactive copilot for building tailored synthetic datasets. 

Together, these copilots tackle challenges that have long constrained scientific and industrial progress: working with noisy or incomplete data, uncovering causal relationships, and creating trustworthy synthetic datasets. By assisting humans in these demanding areas, copilots represent a powerful new model of collaboration: one where AI extends human capability without replacing human judgment. 

Each of these pillars represents not just a specific application, but a rethinking of what agents can and should do. Together, they show the breadth of what’s possible when AI becomes an active collaborator. 

The Future: Beyond Single Agents 

If the past has been about laying foundations, and the present is about consolidating pillars, the future of AI agents will be defined by moving from advanced powerful individual agents to networks of agents working together. 

Individual agents – whether they are digital twins, planners, copilots, or autoformalism systems – already demonstrate remarkable capabilities. But their true potential lies not in isolation, but in interaction. When agents collaborate, they can exchange information, specialise, and coordinate strategies far beyond the limits of any one system.  

This is where our Genies framework comes in. Introduced earlier this year, Genies distil the lessons of reality-centric AI into a vision of the next generation of agents. Not a single-purpose assistants but multi-capable companions that can integrate knowledge, reason about goals and strategies, and adapt continuously. Unlike today’s agents, Genies are designed with eight foundational pillars of intelligence – including creativity and lifelong learning – that make them proactive collaborators rather than reactive tools. Crucially, their purpose is not replacement but human empowerment: Genies are built to amplify human capability, helping people learn, reason and discover at a faster pace. They embody what an individual agent of the future should look like. 

Building on this, The Agent Network sets out a complementary vision: what happens when such advanced agents interact. In an agent network, different agents each with distinct strengths exchange information, specialise, and coordinate strategies. A digital twin agent might simulate patient outcomes, feeding insights to a planning agent, which collaborates with an autoformalism agent to translate findings into new hypotheses, all overseen by a copilot ensuring human experts remain in control. Beyond collaboration, these networks also open space for competition, negotiation, and communication; dynamics that mirror human societies.  

Together, Genies and the Agent Network redefine the future of AI. Genies chart the path for what an agent can be; the Agent Network envisions how agents will work together. This shift transforms AI from isolated tools into a societal infrastructure for intelligence, an ecosystem of partners interacting with each other and with us. 

In our lab, we see this as the next great frontier of AI research. Our past work provides the building blocks. The future will be about how advanced agents evolve, how networks interlock and how they can be trusted to act responsibly in high-stakes domains. By envisioning and prototyping this future now, we aim not just to follow the direction of AI, but to help shape it.