Agents have burst onto the scene in the last year. Agentic AI refers to AI systems that can pursue a goal, make decisions, take actions, and then adapt based on the results. 

Unlike traditional AI models that mostly answer questions or classify information, an agentic system can: 

• break a task into steps;
• run experiments or simulations;
• evaluate outcomes; &
• decide what to do next.

This makes Agentic AI particularly powerful in areas where autonomy and multiple step tasks are required. This includes across research, engineering and innovation, where rapid iteration and autonomous exploration can significantly speed up discovery. 

Agentic systems act autonomously, so risks extend beyond model outputs and with higher potential harms. Challenges include: 

• ensuring safety and reliability of autonomous decisions;
• preventing “hallucinated” or unsafe actions;
• coordinating oversight and accountability; &
• managing cyber-physical interactions in labs, factories and robotics.

Despite all of the recent attention paid to LLM-based agentic frameworks, it’s worth bearing in mind the decades of work of Prof. Michael Wooldridge, who has been working on agents since the late 1980s:

Large language models miss the multi-agent mark La Malfa, G. L. M., Zhang, J., Black, E., Luck, M., Marro, S., La Malfa, E., Wooldridge, M. & Torr, P., 2025, NeurIPS 2025 – Position Track.