AI Agents in Action Foundations for Evaluation and Governance 2025

Looking ahead: multi- agent ecosystems3 Future ecosystems of interacting agents introduce new risks that demand interoperable standards and oversight. Future ecosystems of interacting agents introduce new risks that demand interoperable standards and oversight. The future of AI agents will happen in a much broader space than enterprise automation and will increasingly be defined by the emergence of multi-agent ecosystems. In these ecosystems, agents are expected to interact, negotiate and collaborate across organizational and technical boundaries. In many ways, the interconnectedness of these systems will redefine the future of AI, moving beyond traditional enterprise automation to allow agents to negotiate, collaborate and coordinate autonomously. While this shift opens new opportunities for innovation, it also introduces challenges around alignment, trust, emergent behaviours and system design. Given the complex nature of these systems, ensuring responsible behaviour and effective use requires robust mechanisms for monitoring and assessing agent interactions. A few examples of emerging multi- agent ecosystems and their implications are: –Agent-to-agent commerce: Agents can initiate transactions, request services or exchange data with other agents, forming a new layer of internet activity with considerable downstream economic implications. –Internet of agents: Beyond isolated interactions, large-scale networks of agents could form an “internet of agents,” raising questions of interoperability, standards, governance and societal impact. –Trust frameworks for inter-agent collaboration: As agents begin operating autonomously across boundaries, establishing shared norms, credentialing systems and behavioural standards is critical to verify identity, capabilities and reliability. –Agent governance and oversight: As agent capabilities advance, dedicated “governor” or “auditor” agents will monitor, audit or regulate the actions of other agents, validating transactions, detecting anomalies and correcting unsafe or unintended behaviours. They enable scalable oversight in complex ecosystems, but they risk overreliance on agents supervising other agents. –Embodied agents: Embodied agents extend governance challenges into the physical world, where oversight mechanisms must address both digital actions and consider physical safety, reliability and human interaction. As organizations begin to deploy multiple agents across departments, systems and networks, a new class of failure modes is emerging, linked to potentially misaligned interactions between agents. A few examples include: –Orchestration drift: When agents are plugged into other agents without shared context or coordination logic, workflows can become brittle or unpredictable. –Semantic misalignment: When two agents interpret the same instruction differently, it can lead to conflicting actions or duplicated effort, with implications for safety, reliability and coordination. –Security and trust gaps: Without shared trust frameworks, agents may inadvertently expose sensitive data or interact with malicious actors, exploiting vulnerabilities in the system. –Interconnectedness and cascading effects: Failures in tightly linked agents or systems can propagate across networks, creating a chain of disruptions. –Systemic complexity: As the number and diversity of interacting agents grow, the likelihood of emergent behaviours and cascading failures increases, making them more difficult to anticipate, trace or diagnose. Although the widespread deployment of multi-agent ecosystems is still in its early stages, providers and adopters must now anticipate the associated risks. As organizations experiment and pilot agents, misaligned interactions are already creating new failure modes. Understanding possible challenges such as orchestration drift, semantic misalignment and cascading failures enables adopters to implement safeguards before scaling. A proactive approach ensures responsible growth, aligning governance with technical capabilities and defined boundaries. As organizations begin to deploy multiple agents across departments, systems and networks, a new class of failure modes is emerging. AI Agents in Action: Foundations for Evaluation and Governance 28