Exploring Agent-Based Computing

The Agentcities initiative operated from 2000 to approximately 2005. The Task Force and its working groups are no longer active, and the testbed that once connected platforms across more than twenty countries has long since been retired. But the intellectual legacy — the research questions it framed, the architectural principles it established, and the hard-won lessons from operating a real global agent network — remains directly relevant to anyone working on autonomous, distributed, or AI-based computing systems today.

Continuing the Research Tradition

The core problems Agentcities addressed have never been more relevant than they are today. Autonomous AI systems — large language model agents, robotic platforms, and distributed decision-making networks — face the same fundamental questions about interoperability, trust, and coordination that Agentcities grappled with two decades ago. The vocabulary has shifted: from FIPA-ACL to natural language instructions, from formal ontologies to vector embeddings, from agent directories to model context protocols. Yet the conceptual challenges are strikingly similar. Agent architectures that were theoretical exercises in 2003 are production realities in 2025, which is precisely why revisiting the Agentcities record repays the effort.

Key Entry Points for MAS Research

For readers who want to engage with multi-agent systems as a living field, the following venues, platforms, and standards bodies are the most direct entry points. Each carries forward, in its own way, a strand of the work Agentcities helped advance.

Historical Documentation

The most complete historical documentation of Agentcities work exists in conference proceedings. AAMAS 2002, 2003, and 2004 papers include substantial Agentcities contributions, and the ECAI (European Conference on Artificial Intelligence) proceedings from the same period document the European side of the participation. The Internet Archive's Wayback Machine preserves snapshots of the original Agentcities.org website, including working group reports, technical recommendations, and event documentation — an invaluable primary source for anyone reconstructing the initiative's timeline.

Modern MAS Platforms

For practitioners who want to work with agent-based systems today, several platforms continue the tradition in different ways. Mesa (Python) supports agent-based simulation and modeling. NetLogo, widely used in education and research, offers visual simulation environments for exploring emergent behavior. SPADE is a modern Python platform with FIPA-ACL support, positioning itself explicitly as a successor to JADE for contemporary deployments. And AutoGen and LangGraph address the emerging field of LLM agent orchestration — building systems in which multiple AI models collaborate — a domain that inherits many of the coordination challenges Agentcities studied long before large language models existed.

Open Testbeds and the Agentcities Model

One of Agentcities' most durable insights was deceptively simple: real progress in distributed systems requires real shared infrastructure. The initiative's decision to build an actual global testbed — not merely to specify protocols on paper — accelerated learning dramatically, surfacing interoperability failures that no amount of specification review could have anticipated. That model continues in contemporary AI research through shared benchmarks, open evaluation frameworks, and multi-organization research collaborations. Whether in reinforcement learning, large language models, or robotic systems, the value of open shared testbeds that Agentcities demonstrated remains a guiding principle for the field.