Guiding AI: A Framework for Responsible Agentic Systems

by

Author: Denis Avetisyan

Researchers have developed a new multi-agent framework designed to embed ethical considerations, sustainability goals, and legal compliance directly into the core logic of autonomous AI systems.

The multi-agent framework streamlines decision-making, translating user input into a finalized, optimized action through a structured flow.
The multi-agent framework streamlines decision-making, translating user input into a finalized, optimized action through a structured flow.

COMPASS integrates value-aligned reasoning, carbon-aware computing, and Retrieval-Augmented Generation to promote responsible AI deployment and digital sovereignty.

The increasing autonomy of large language model-based agents presents a critical challenge: ensuring alignment with complex societal values. This paper introduces ‘COMPASS: The explainable agentic framework for Sovereignty, Sustainability, Compliance, and Ethics’, a novel multi-agent system designed to systematically integrate digital sovereignty, environmental sustainability, regulatory compliance, and ethical considerations into autonomous decision-making. By employing a \text{LLM\text-as-a-judge} methodology and Retrieval-Augmented Generation, COMPASS provides both quantitative assessments and explainable justifications for its evaluations. Will this composition-based approach facilitate the responsible deployment of agentic AI across diverse and evolving application domains?

The Evolving Landscape of Agentic AI

The emergence of agentic AI signifies a shift towards autonomous systems capable of defining and executing tasks with minimal human intervention, largely driven by advancements in large language model (LLM)-based agents. These agents demonstrate the potential to revolutionize problem-solving across diverse domains, from complex data analysis to automated scientific discovery. However, realizing this potential is contingent on overcoming significant hurdles related to alignment and trustworthiness. Ensuring these AI systems consistently pursue intended goals-and avoid unintended consequences-requires robust mechanisms for specifying desired behavior and verifying outcomes. Furthermore, establishing trust necessitates addressing concerns about potential biases, unpredictable actions, and the difficulty of attributing responsibility when autonomous agents operate in complex, real-world scenarios. Without resolving these challenges, the promise of agentic AI risks being overshadowed by legitimate anxieties regarding safety, reliability, and ethical implications.

The pursuit of truly agentic artificial intelligence is currently hampered by inherent limitations within large language models (LLMs). While capable of impressive feats of text generation, LLMs often struggle with tasks demanding deep, multi-step reasoning or consistently accurate factual recall – critical components for reliable autonomous action. Simple scaling of model size isn’t proving sufficient to overcome these deficiencies, necessitating exploration of novel architectural approaches. Researchers are investigating methods like incorporating symbolic reasoning modules, knowledge retrieval mechanisms, and improved methods for verifying information to augment LLM capabilities. These advancements aim to move beyond pattern recognition and towards genuine understanding, ultimately fostering AI systems that can not only act independently, but also do so with a degree of trustworthiness and intellectual rigor.

Realizing the transformative potential of truly autonomous AI demands more than just technical refinement; it necessitates a proactive assessment of wider societal consequences. While overcoming current limitations in reasoning and factual accuracy is paramount for building reliable agentic systems, such progress inevitably raises complex ethical and practical challenges. Considerations extend beyond algorithmic bias to encompass potential job displacement, the responsible deployment of autonomous decision-making, and the safeguarding of human control over critical infrastructure. Ignoring these broader impacts risks eroding public trust and hindering the beneficial integration of agentic AI into daily life, ultimately requiring a collaborative approach between researchers, policymakers, and the public to ensure equitable and sustainable development.

Orchestrating Intelligence: The COMPASS Framework

The COMPASS Framework utilizes a multi-agent system architecture to govern the behavior of autonomous AI agents. This orchestration is specifically designed to ensure alignment with four core principles: digital sovereignty, promoting user control over data; sustainability, minimizing environmental impact; compliance, adhering to relevant regulations and standards; and ethics, embedding moral considerations into agent decision-making processes. By distributing tasks across multiple specialized agents and implementing centralized oversight, COMPASS aims to mitigate risks associated with unaligned AI and promote responsible development and deployment of agentic systems. The framework’s architecture facilitates the integration of governance mechanisms at each stage of an agent’s operation, from initial planning and knowledge retrieval to action execution and outcome evaluation.

The COMPASS framework utilizes Retrieval-Augmented Generation (RAG) to enhance agentic AI performance by grounding responses in verified knowledge sources, rather than relying solely on the LLM’s pre-trained parameters. This process involves retrieving relevant documents based on user input and incorporating them into the prompt, providing agents with contextual information to improve accuracy and reduce hallucination. Furthermore, COMPASS employs a Large Language Model as a Judge (LLM-as-Judge) to objectively evaluate agent outputs, assessing alignment with defined criteria and ensuring responsible decision-making. Quantitative analysis, specifically using the BERTScore metric, demonstrates a measurable improvement in semantic coherence resulting from the implementation of these techniques, indicating that RAG and LLM-as-Judge contribute to more logically consistent and factually grounded agent actions.

The Synchronizing Agent functions as a central structural element within the COMPASS framework, enabling the creation of specialized agents through inheritance and customization. This agent facilitates the integration of retrieved information into the decision-making process of other agents, and performance evaluations, quantified by ΔScore metrics, demonstrate a measurable positive impact of this retrieved information on agent output. Specifically, ΔScore represents the difference in performance between agent actions with and without the inclusion of externally sourced, verified knowledge, indicating the Synchronizing Agent’s efficacy in grounding agent behavior and improving overall system coherence.

Establishing Trust: Accountability and Ecological Responsibility

The Ethos Blockchain integrates with the COMPASS framework to establish a verifiable record of all actions performed by autonomous agents. This is achieved through the creation of an immutable audit trail, where each agentic action is recorded as a transaction on the blockchain. This functionality enables post-hoc accountability; any action can be traced back to its origin and verified for compliance with pre-defined protocols or agreements. The resulting transparency fosters trust in the agent’s behavior and provides a mechanism for dispute resolution or error correction, as all actions are permanently and publicly recorded on a distributed ledger.

COMPASS incorporates Carbon-Aware Computing principles to mitigate the environmental impact of AI workloads. This is achieved through direct integration with tools such as CodeCarbon, which enables the tracking, reporting, and ultimately, the reduction of carbon emissions associated with model training and inference. By quantifying the carbon footprint of AI processes, COMPASS facilitates informed decision-making regarding resource allocation and model optimization, allowing developers to prioritize energy efficiency and minimize their contribution to climate change. This integration provides actionable data for reducing the environmental cost of AI operations.

The growing emphasis on sustainable artificial intelligence is actively supported by organizations such as the Green AI Institute. This institute focuses on research and advocacy for practices that reduce the environmental impact of AI systems, specifically addressing the substantial energy consumption associated with model training and deployment. Their work highlights the necessity of quantifying and minimizing the carbon footprint of AI, promoting techniques like efficient model design, optimized hardware utilization, and the use of renewable energy sources to power AI infrastructure. This external validation underscores the commitment of initiatives like COMPASS to environmentally responsible AI development and operation.

Safeguarding Sovereignty: AI Governance for a Decentralized Future

The COMPASS framework is intentionally designed to bolster Digital Sovereignty, recognizing that artificial intelligence systems operate within, and should reflect, the distinct legal and ethical landscapes of individual regions. This commitment moves beyond simple compliance, actively ensuring AI aligns with locally-defined values and regulations, rather than being dictated by a handful of global technology providers. A prime example of this support is COMPASS’s compatibility with initiatives like Gaia-X, a European project aiming to create a secure and trustworthy data infrastructure that prioritizes user control and data portability. By embracing such frameworks, COMPASS fosters a future where AI innovation is guided by regional priorities, strengthening local economies and empowering citizens with greater control over their digital lives.

The architecture of COMPASS is fundamentally built upon existing and emerging legal frameworks designed to govern artificial intelligence. Specifically, alignment with the European Union’s proposed Artificial Intelligence and Data Act (AIDA) ensures responsible data handling and algorithmic transparency, while the guiding principles of the Monteal Declaration for Responsible AI – encompassing human rights, environmental sustainability, and democratic values – are woven into its core functionalities. This commitment isn’t merely about compliance; it represents a proactive approach to building trustworthy AI systems, embedding ethical considerations from the outset to mitigate potential risks and foster public confidence. By prioritizing these established guidelines, COMPASS aims to establish a standard for AI development that is both innovative and ethically sound, paving the way for responsible technological advancement.

The COMPASS initiative envisions a future where artificial intelligence serves as a powerful catalyst for individual empowerment and societal fortitude, achieved through a deliberate emphasis on ethical frameworks, environmental sustainability, and regulatory compliance. This approach transcends mere technological advancement; it proactively addresses potential harms and ensures AI systems are developed and deployed responsibly. By integrating these principles into the core design, COMPASS aims to foster public trust and enable widespread adoption of AI technologies that genuinely benefit communities, bolster resilience against unforeseen challenges, and ultimately contribute to a more equitable and sustainable future for all.

The presented framework, COMPASS, endeavors to impose structure upon the inherent chaos of agentic AI. It prioritizes demonstrable alignment with ethical and legal boundaries, a necessary curtailment of potential computational drift. This echoes John von Neumann’s assertion: “The sciences do not try to explain why we exist, but how we exist.” COMPASS doesn’t attempt to define morality itself, but rather to establish a functional, verifiable mechanism for its implementation within complex systems. The focus on sustainability and compliance isn’t idealistic; it’s a pragmatic acknowledgement that even the most sophisticated intelligence requires defined parameters to avoid unintended consequences, mirroring the rigorous constraints within which all effective computation operates.

What’s Next?

The pursuit of agentic AI, predictably, has yielded more questions than resolutions. COMPASS offers a structural response-a framework for embedding ethical constraints-but the devil, as always, resides in the granular detail of those constraints. The elegance of the proposed system lies in its attempt to reduce complexity, to articulate principles as operational logic. Yet, one suspects the true test will not be the framework itself, but the sheer difficulty of achieving consensus on the values it embodies. A system is only as virtuous as its axioms.

Future iterations must confront the inherent limitations of ‘value-alignment’. Can a static ethical framework truly account for the dynamic, unpredictable nature of real-world scenarios? The reliance on LLM-as-Judge, while pragmatically sound, introduces the well-documented biases inherent in large language models. Mitigating these biases-and acknowledging their irreducible presence-remains a critical challenge. The focus should shift from building more complex systems to refining the precision of their underlying principles.

Ultimately, the success of frameworks like COMPASS will be measured not by their technical sophistication, but by their capacity for graceful failure. A truly robust system doesn’t strive for perfect compliance; it anticipates deviations and minimizes harm. Perhaps the most fruitful avenue for future research lies not in perfecting the ‘agent’, but in designing the mechanisms for responsible disengagement.

Original article: https://arxiv.org/pdf/2603.11277.pdf

Contact the author: https://www.linkedin.com/in/avetisyan/

See also:

2026-03-14 16:26