Human-centred collaboratIVE MultI-ageNt framework for accelerating software Development and maintenance (HIVEMIND)

Overview

Context:

Software engineering is changing fast, yet development cycles are lengthening, systems are growing more complex, security risks are rising, and maintenance remains a persistent challenge. Existing AI tools help, but many lack adaptability, contextual understanding, and effective human oversight, often producing unreliable outputs that require manual correction.

Project objective:

HIVEMIND closes these gaps with a responsible, human-centered, multi-agent LLM framework that improves collaboration between developers and AI agents across eight specialized roles: ElicitAgent, ArchitectureAgent, TaskGenAgent, CodeGenAgent, TestGenAgent, QualityAgent, DeploymentAgent, and MaintainAgent. The framework supports human-to-agent and agent-to-agent interaction and achieves multi-layered specialization through prompt engineering, fine-tuning, retrieval-augmented generation, and reinforcement learning from human feedback. Additionally, HIVEMIND builds and validates intelligent software for five mission-critical use cases: citizen engagement, disaster response, healthcare, automated mobility, and manufacturing, thus achieving Technology Readiness Level 5.

Our role:

IDeA-IDSIA leads the design and implementation of the multi-agent collaborative framework for responsible software engineering. We define interaction mechanisms, protocols, and APIs that enable seamless agent collaboration, and we develop validation and testing processes while continuously monitoring emerging multi-agent AI approaches to incorporate useful innovations. Our ecosystem forms the backbone of HIVEMIND, enabling customization of multi-modal LLMs into specialist AI agents and defining their behavioral patterns for software engineering tasks.

Publications

At the moment, we have two publications under review. The work on the third publication is in progress.