AI-assisted software development is moving quickly, but HIVEMIND is looking beyond individual tools and short-lived technological trends. Halfway through the project, our partners describe a broader ambition for a modular multi-agent environment that can adapt as AI evolves, work with an organisation’s own knowledge, keep humans involved in critical decisions and remain trustworthy enough to be used in real development workflows.
In our mid-term interview series, researchers from five HIVEMIND partner organisations reflected on what makes this approach different, the challenges they are encountering and what they have learned while turning the project’s ideas into working technologies.
Moving beyond isolated AI tools
For Santiago Massé and Eduardo Arnesto from IDENER, one of HIVEMIND’s main differences lies in how AI models are connected with established software engineering practices. Rather than focusing only on model performance, the project is developing an environment where organisations can use models trained on their own data and create agents with knowledge relevant to their particular context.
This approach also has implications for privacy and adoption. Models can be deployed within an organisation’s own infrastructure, while the framework is designed to fit existing workflows rather than requiring teams to completely change the way they work. At IDENER, this has included work on coding agents and systems that allow agents to access external information that was not part of their original training.
Keeping humans involved when AI makes decisions
At the Universidad de Alicante, Alba Bonet Jover looks at the same challenge from the human side of the interaction. As she puts it, “Developers need to know how AI reaches decisions.” Good outputs alone are not enough to build trust. People also need ways to understand, assess and influence what AI systems are doing.
The Alicante team has therefore focused on human-computer interaction and feedback methodologies that keep experts involved throughout the software development lifecycle. One of their main achievements has been the development of guidelines for interaction with the HIVEMIND environment, adapted to the different agents being created within the project. The objective is to find a workable balance between automation and human judgement rather than treating the two as alternatives.
Building a system that can continue to evolve
Mobyen Uddin Ahmed from Mälardalen University describes HIVEMIND as “not only a static architecture, but a dynamic and human-centred system”. That distinction matters because the AI landscape around the project is changing continuously, with new models, techniques and expectations around responsible AI emerging all the time.
HIVEMIND is being designed with this evolution in mind, using continuous evaluation, updated knowledge and mechanisms that allow individual capabilities to improve over time. Mälardalen University is also contributing responsible AI principles to the project, including explainability, fairness, accountability and human oversight. The goal is to make sure that greater technical capability develops alongside the safeguards needed for people to understand and trust the system.
Designing for change rather than predicting it
Marc Oriol from Universitat Politècnica de Catalunya approaches adaptability from the perspective of software architecture. “It’s not about trying to predict where AI is going. It’s more about how we can adapt to it.”
That idea is reflected in HIVEMIND’s modular and open-source design. New technologies can be incorporated without rebuilding the entire environment, while organisations can select the agents that are most relevant to their own needs instead of adopting the full system at once.
Marc also emphasises how much the performance of AI agents depends on the quality of the data behind them. For UPC, one of the biggest challenges has been finding and preparing suitable data for agents working on requirements elicitation, software architecture and task generation. It is a reminder that useful AI depends not only on models and architectures, but also on the less visible work of building reliable foundations for them.
Trust becomes even more important when agents work together
The final interview in the series turns to what happens when multiple AI agents begin to depend on one another. Mazen Mohamad and Robert Lowe from RISE argue that agentic AI can help accelerate software development, but only if that happens in a way that remains trustworthy and transparent.
HIVEMIND uses specialised agents across different phases of the software development lifecycle, extending beyond coding and testing to areas such as deployment, maintenance, quality and security. This also means that the appropriate level of human supervision can vary depending on the task. A security-related decision, for example, may require much closer oversight than a more routine automated process.
At RISE, part of this work focuses on agents that assess software quality and identify cybersecurity risks, including known vulnerabilities. The team is also looking at a challenge that becomes particularly important in multi-agent environments. If one agent produces an incorrect or hallucinatory output and another agent then uses it as input, the problem can spread through the system. Detecting and mitigating these failures is therefore an important part of making collaborative AI agents dependable in practice.
One project, five views of the same challenge
The five interviews point towards a common direction for HIVEMIND. The project is not built around the idea of one powerful AI assistant working in isolation, but around specialised agents that can collaborate, adapt to new technologies, use organisation-specific knowledge and work alongside human experts across the software development lifecycle.
Getting there involves more than improving AI performance. It also means addressing interoperability, data quality, transparency, security, human oversight and the practical question of how these technologies can fit into the organisations that may eventually use them.
Each partner approaches those questions from a different part of the system. Together, their work is gradually turning HIVEMIND from a multi-agent concept into an environment that can be tested against the realities of software development.