The reliable application of LLM-based agents to SE requires a tremendous increase in their accuracy and minimisation of their bias. While LLMs continue increasing in size and performance, it seems that phenomena like hallucinations of a single agent are substantially inevitable, since they are linked to the fundamental inference mechanism in generative models. On the other hand, evidence starts accumulating about the possibility of achieving the required performance by collaboration and debate among groups of agents. As it happens among humans, quality of work increases with specialisation of workers on tasks, organised collaboration, and discussion among workers with different backgrounds. Differently from humans, the instantiation of multiple required AI agents, and the collaboration and discussion among them, are very fast and cheap, making this approach particularly convenient. MOSAICO proposes the theoretical and technical framework to implement this approach and to scale it to very large groups of collaborating agents, i.e. AI-agent communities. The developed solutions are composed into an integrated MOSAICO platform, handling communication, orchestration, governance, quality assessment, benchmarking and reuse of AI agents. MOSAICO is integrated with existing development environments, to present the results to software engineers, and allow expert users to intervene in the AI decisions. The performance and reliability of MOSAICO technologies and tools to achieve given software engineering tasks are assessed within 4 different use cases scenarios coming from immersive technologies, bank/financing, aerospace and Internet of Things sectors. The long-term adoption of MOSAICO results and technologies will be ensured by open sourcing the code and fostering an open collaboration, such as open-source initiatives, to enhance user engagement in the MOSAICO community.