Agentic AI needs a science of interaction

Manlio De Domenico 
Nature Communications (2026)

Artificial intelligence is moving from stand-alone models towards systems of interacting agents. These systems do not merely classify, predict or generate: they plan, exchange messages, call tools, divide labour, adapt and coordinate action across tasks and time. This shift has made agentic AI one of the dominant frames for the next phase of the field. Understanding how these agents interact will therefore be as important as improving what each agent can do alone.

Read the full article at: www.nature.com

Stigmergy: bottom-up coordination, from molecules to collective intelligence

Francis Heylighen, Evo Busseniers, Orion Maxted, Karin Verelst, Clement Vidal

Stigmergy is a self-organizing mechanism of coordination in which the traces left by actions in a shared medium elicit complementary actions, without need for supervision or direct communication between the agents. Originally proposed to explain the collective activity of social insects, the concept has since been extended to a wide range of domains, from chemical reactions and social organization to individual cognition, AI, consciousness, the Internet, and collective intelligence. We briefly review these applications, showing how stigmergy enables a spontaneous division of labor, workflow, and aggregation of contributions, as illustrated in particular by the development of Wikipedia. To explore its potential at a smaller scale, we organized two demonstration experiments in which about 40 anonymous participants collaboratively wrote a story and a discussion paper. The results exhibited remarkable creativity and conceptual diversity, but limited coherence, suggesting the need for additional time and mechanisms for refinement and integration. We outline directions for further investigating and improving stigmergic collaboration, particularly through metamarks that guide further contributions.

Read the full article at: researchportal.vub.be

Computational models of complex systems meet AI agents

Ciro Cattuto & Alessandro Vespignani 
npj Complexity (2026)

Complex systems have structural properties that make them a promising domain for scientific workflows driven by AI agents— systems that combine large language models with tool use and iterative execution. However, most complex systems simulators and libraries do not expose their control points and outputs in a form that text-based AI agents can readily use. Realizing the potential of agent-driven workflows requires equipping complex systems models with agent-friendly interfaces.

Read the full article at: www.nature.com

The Four Forces That Shape Collective Memory – Nautilus

Physicist César Hidalgo has been studying who becomes famous and what shapes our collective memory of popular figures for over a decade. What makes someone like Napoleon a household name across the world, whereas Bernando O’Higgins, a Chilean military officer who helped to lead that nation’s successful battle for independence, is only recognized within a small corner of South America? How did the rise of printing, television and other technologies shape who we tell stories about and remember for decades to come? 

A pioneer in the field of economic complexity, Hidalgo’s interest arose in part out of his work on a project called the Observatory of Economic Complexity, which uses trade data to map out the commercial goods different countries make and sell to one another. He wondered if there might also be a way to trace the cultural products that countries produce and spread around the globe—and the forces that drive these scientific, intellectual, and artistic currents. So in 2013, he created another online observatory called pantheon.world, a repository of collective cultural memory. 

Read the full article at: nautil.us

Flatness-based control of nonlinear complex networks

I. Sendiña-Nadal, I. Leyva, J.P. Barbot, C. Letellier

Effective control of nonlinear complex networks needs to identify not only the intervention sites but also the state variables to be monitored and the locations at which control inputs should be applied. Here, we present a network-scale generalization of an observability-guided, flatness-based control framework. The resulting control laws are implemented using sliding-mode control and evaluated in heterogeneous networks of chaotic oscillators driven toward a prescribed periodic regime. We compare two control architectures: in the node-wise method, oscillators are independently sensed and actuated, while in the pairwise method, measurements acquired from a single node within each dyad are used to synthesize control inputs applied to both nodes. Across the parameter regimes considered, node-wise control generally achieves accurate control over a broader operating region, whereas pairwise control remains effective over a more restricted region while using half as many measurement locations for a given number of actuated nodes. Moreover, network interactions can enhance control by propagating the control influence to non-actuated nodes, creating a cooperative effect between the collective dynamics and the control action. Collectively, these findings quantify the trade-offs among measurement-location requirements, control performance, and control effort in nonlinear complex networks.

Read the full article at: arxiv.org