Collaborative AI: The open source path (MOSAICO)

"Collaborative AI: The open source path," presented by Massimo Tisi (IMT Atlantique), explores multi-agent architectures and orchestration strategies for improving reliability in AI systems. This workshop session is part of the Open Community for Research at Open Community Experience 2026 in Brussels, Belgium. This session introduces the MOSAICO project, which focuses on orchestrating and supervising communities of AI agents to improve reliability in software engineering tasks. It examines how single-agent approaches struggle with complex specifications, particularly in generating syntactically and semantically correct artefacts such as domain models. A multi-agent approach is presented where tasks are decomposed across specialised agents responsible for generation, validation, and evaluation. These agents can use different models, tools, and configurations, enabling task-specific optimisation, variability through replication, and cost-aware execution strategies. The MOSAICO architecture introduces a hierarchical system of agents, including reference agents, collaboration agents for workflow decomposition, solution agents that execute tasks, and supervision agents that evaluate outputs. Consensus agents resolve conflicting evaluations, supported by orchestration and decision engines coordinating execution. The platform emphasises explicit collaboration structures, agent reputation through benchmarking, and isolation via containerised execution. Agents communicate through defined protocols, enabling interoperability across languages and frameworks while supporting observability and evaluation of agent performance. Key topics covered - multi-agent ai systems - ai agent orchestration - task decomposition in ai workflows -syntactic and semantic validation agents - hierarchical agent architectures -consensus mechanisms in ai systems - agent benchmarking and reputation - orchestration and decision engines - containerised agent execution - interoperability protocols for agents - ai reliability in software engineering - mosaico project architecture Why this matters Multi-agent systems introduce structured approaches to validation, coordination, and evaluation, addressing reliability limitations of single-agent AI workflows in complex engineering tasks. About OCX26 Open Community Experience 2026 is the Eclipse Foundation’s flagship event, held in Brussels, Belgium. It brings together developers, architects, and industry leaders to explore open source technologies across domains including AI, automotive, tooling, and cloud systems, with a focus on practical implementation. Learn more at https://www.ocxconf.org/ Chapters 00:00 introduction and project overview 00:46 example task: generating domain models 01:44 limitations of single-agent approaches 02:16 multi-agent architecture fundamentals 03:18 task delegation and sub-agents 03:53 benefits: specialisation, variability, cost 04:39 challenges: latency and error propagation 05:01 research perspectives on multi-agent systems 06:29 introduction to mosaico project 07:26 mosaico client and vs code integration 08:50 hierarchical agent architecture 09:50 solution, supervision, and consensus agents 11:15 orchestration and decision components 12:03 platform components and delivery 12:35 design principles and architecture choices 13:24 reputation-based agent selection 14:20 explicit collaboration and agent models 15:26 formal semantics and protocols 15:45 open source model and community contributions 16:48 consortium and use cases 18:17 demo environment and deployment 20:12 protocol design and fipa influences 21:30 belief-desire-intention (bdi) usage