AI Lab: Survey of SoTA in Agentic LLMs

Multi-agent systems (MAS) have been researched for decades, but the arrival of LLMs gave the field a huge boost, as prior MAS systems could in practice only be deployed in very well defined domains.

And on the flip side, agentic approaches to LLMs and LLM-powered agents have turned LLMs from passive couch potatoes, into active AI systems that can get work done. For a discussion of this evolution, see my PegaWorld breakouts from 2024 and 2025 here, and there will be a lot more exciting agentic announcements at PegaWorld this year:

So with so much going on, how to keep up with the more researchy view on LLM-powered agents? To address this need I collaborated with my Leiden University colleagues on the survey paper below. This was mostly written at the start of 2025 and published at the end of year, which might feel like eons ago in agentic AI, but it should still give you loads of pointers where agentic research might be heading!

Aske Plaat, Max van Duijn, Niki van Stein, Mike Preuss, Peter van der Putten and Kees Joost Batenburg. Agentic Large Language Models, a survey. Journal of Artificial Intelligence Research, Vol. 84, Article 29. Publication date: December 2025.

What agentic topics or reseach developments should be added in a 2026 overview do you think?

Enjoy watching and reading!

Thanks for sharing this, @Peter_van_der_Putten

it reinforces two points that can’t be emphasized enough.

  1. agentic capability is fundamentally an architectural concern, not something that can be solved through prompting alone.
  2. as agents gain the ability to act in the real world, safety, governance, and control become first‑order design requirements, not afterthoughts

Addressing this properly comes back to sound solution architecture:

  1. Clear business intent, contextual requirements, and disciplined design choices
  2. While GenAI gives developers speed and power to build and “vibe‑code” solutions, those artefacts still need to conform to architectural and design guardrails. This is where human review, experience, and engineering craftsmanship matter more than ever.