The first year of the HIVEMIND project has drawn to a close, and the consortium can now look back at twelve months of intensive development, coordination and conceptual work. As Project Coordinator Daniel García Guirao notes, “HIVEMIND aims to show how coordinated AI agents can amplify human expertise, not replace it.” Although the project is still at an early stage, the groundwork laid during this period will influence every technical, scientific and ethical step that follows. HIVEMIND stands at the intersection of large-scale AI engineering, human-centred design and a commitment to trustworthy and accountable systems, and the first year has been dedicated to ensuring that these elements grow together in a coherent way.
The past year has also been a time of discovery within the consortium itself. Each partner has contributed from different disciplinary backgrounds, and this diversity has helped establish a shared direction for the project’s technical architecture, user requirements and governance principles. Their reflections reveal how the ecosystem is beginning to take form and where the priorities lie as HIVEMIND moves into its next phase.
Setting the technical groundwork
IDENER, represented by Project Coordinator Daniel García Guirao and Technical Coordinator Santiago Masse, opened the project year by focusing on the structural and organisational elements that a multi-agent ecosystem demands. They consider the successful launch of the project and the preservation of an interdisciplinary approach as essential achievements, along with the consolidation of a comprehensive set of technical requirements. The definition of the overall system architecture and the creation of an initial framework for the specialisation of language models have set the technical baseline on which the coming stages will rely.
Looking ahead, IDENER places its attention on designing and deploying the agents that will constitute the HIVEMIND ecosystem. They note the need to fulfil ethical, technical and user requirements in parallel, ensuring that the system provides real scientific and societal value. Maintaining strong integration across technical work packages will also be a priority, so that the ecosystem can scale and operate seamlessly.
Keeping human judgement at the centre
Alba Bonet from the University of Alicante shared that they have been working on the Human-Computer Interaction (HCI) guideline that will help define how users engage with the agents within HIVEMIND. The team began drafting the first version of the guideline during the project’s first year, identifying fields that will be relevant across different agents and scenarios. An initial meeting with IDENER helped align the design of interaction and feedback mechanisms with the broader system architecture.
Bonet explains why this step matters for the project’s responsible AI ambitions. Human judgement, she argues, must remain central to the framework, as it reduces risks linked to ethical concerns, hallucinations and bias. The guideline aims to support human-in-the-loop strategies that encourage effective collaboration between users and agents. As she puts it, the guideline is intended to “facilitate effective interaction between humans and agents, ensuring adequate human oversight in tasks that demand stronger human involvement.”
The team will now work on adapting the general guideline to the specific agents and use cases defined in HIVEMIND, in collaboration with other consortium partners. This process will refine how feedback mechanisms operate and how oversight remains meaningful in practice. Alba Bonet reflects that HIVEMIND has the potential to “find the right balance between automation and human judgement,” and sees the guideline as a key tool in achieving this.
Ethical intelligence through agent design
Professor Mobyen Uddin Ahmed explained that the Mälardalen University focused its first-year effort on understanding how responsibility and trustworthiness can be operationalised within multi-agent AI systems. The team carried out a comparative analysis of existing agent frameworks and examined them through the lenses of transparency, interpretability, bias and accountability. This review revealed persistent gaps in current systems, particularly in the absence of mechanisms for monitoring bias, providing ethical oversight or ensuring fairness in complex workflows.
To address these gaps, the team proposed a conceptual architecture for an “ethical agent,” designed to identify and mitigate misalignments with human values. This architecture incorporates moral alignment techniques, reinforcement learning from human feedback and direct preference optimisation. The team also drafted a traceability and accountability mechanism that makes it possible to track decisions or errors back to their source within the agent ecosystem.
Mobyen Uddin Ahmed summarises the importance of this work through a short statement: “Responsible AI in HIVEMIND means AI that not only works well but behaves ethically—transparent in its reasoning, fair in its decisions, and accountable to human oversight.”
The next phase of the team’s work will translate these early concepts into practical components that can be tested within HIVEMIND’s real-world use cases, contributing to the Ethics Guidelines for Responsible LLM-based Multi-agent Ecosystems.
Understanding user requirements across the software lifecycle
HAVELSAN’s team focused their first year on clarifying user requirements and developing a shared understanding of the software development lifecycle relevant to the project. Through a series of structured discussions and internal analyses, the team refined operational needs and strengthened its alignment regarding the lifecycle elements it contributes to. These preparatory steps will make it easier to test and adapt software tools as HIVEMIND evolves.
The team now anticipates the first iterations of the SDLC tools developed within the project and is preparing to begin systematic testing and feedback processes during the second year.
Data foundations and preparation for validation
At DFKI, the team’s principal achievement during the first year centred on ensuring that data transfer processes were completed for all use case partners. This transfer marks a necessary step for the development and testing phases that follow. The team will now move towards building the validation test suite for the agents, a component that will be essential for assessing how well the system performs under real conditions.
During the first year, the UPC team consolidated the group and planned upcoming tasks They also published two papers, “What About Emotions? Guiding Fine-Grained Emotion Extraction from Mobile App Reviews” and “Multi-Agent Debate Strategies to Enhance Requirements Engineering with Large Language Models” (both available in the Resources section of our web), which reflect their broader research activity relevant to HIVEMIND. Looking ahead, the team identifies the development of agents that satisfy the use case requirements as their principal contribution for the coming year.
Ready for the second project year
HIVEMIND’s first year has been defined by cross-team coordination, foundational design and an ongoing commitment to responsible and human-centred AI development. Although the agents themselves will take shape during the second year, the conceptual and architectural work completed so far ensures that upcoming development rests on a coherent base.
As the project advances, attention will turn to operational prototypes, practical validation and the creation of ethical, transparent and accountable agent interactions. The consortium enters its second year with shared structures, defined responsibilities and a clearer sense of how a multi-agent ecosystem can support robust scientific and societal outcomes.