Laboratorio de Ideas
How conversations around sovereignty are redrawing the digital map
Artículo
24 nov 2025
By Richard Baird, CTO for Core Enterprise & zCloud, and Guilherme Cartier, Associate Director for Infrastructure/Cloud Architecture at Kyndryl
The rise of agentic AI systems marks a significant evolution in artificial intelligence, shifting from traditional rule-based systems to dynamic, decision-making entities that adapt and learn in real-time. These systems are poised to revolutionize industries, including mainframe technology, by enabling more personalized, efficient and autonomous operations.
The future of agentic AI is marked by the emergence of AI agents, the adoption of which is driven by a need to manage complex tasks autonomously and respond to ever-changing environments more efficiently.
What makes agentic AI more than just another wave of automation is that it forces enterprises to rethink longstanding assumptions about how technology is deployed and how processes are implemented. For decades, IT strategy has largely been about codifying best practices into predictable processes; success came from reducing variability. Agentic systems flip that logic. They can thrive in ambiguity and adapt on the fly. The organizations that win in this era won’t be those that simply plug agents into old workflows, but those willing to redesign their operating models around flexibility, resilience and human-machine collaboration.
The architecture behind agentic AI
Agentic systems are built on large language models (LLMs) augmented with capabilities such as retrieval, tools and memory, and they can be categorized into two distinct architectures.
The first are workflow-driven systems, which follow predefined code paths and orchestrate LLMs and tools to deliver predictable, consistent results. They work best for tasks that can be broken down into fixed subtasks, such as automated reporting or data validation.
The second are agents, which dynamically determine their own processes and tool usage, retaining control over how they achieve outcomes, but always keeping a human in the loop to approve or decline potentially disruptive actions. This flexibility makes them well-suited for more complex and fluid scenarios, ranging from customer support to resource allocation.
Kyndryl launched the Kyndryl Agentic AI Framework to support the development of agentic AI systems. This Framework orchestrates self-directed, self-learning agents that work with people to execute complex workflows at scale. It integrates with the existing environment, leverages Kyndryl Bridge where appropriate, and uses graph-based orchestration to balance speed, resilience and transparency.
Complementing these frameworks, the Model Context Protocol (MCP) has emerged as an open standard that simplifies integration of AI models with diverse tools and data sources. By standardizing interoperability, MCP enables developers to build agents that can move more easily across frameworks without extensive reimplementation. This improves flexibility, scalability and maintainability, ultimately streamlining the development of agentic AI.
Practical application in mainframe environments
By designing agents that can carry out specific tasks, such as executing automated maintenance routines or managing resource allocation, organizations can streamline operations and reduce the burden on human operators. These agents can be programmed to perform routine tasks with precision and consistency, bolstering the reliability and stability of the mainframe systems.
Agentic systems can also be designed as intelligent assistants, equipped with deep knowledge about specific organizations and contextual understanding of best practices and procedures. By drawing on this embedded expertise, they can help manage them more efficiently and with greater confidence.
In customer service, for example, agents could provide insights into account management, recommend actions based on historical data, and drive adherence to established protocols.
They can also transform training and onboarding by offering interactive, personalized learning experiences that adapt to the learner’s pace and expertise level. Through real-time feedback, practical simulations, and progress tracking, new hires are better prepared to acquire the skills they need quickly and effectively.
Beyond these day-to-day applications, agentic systems hold promise in large-scale modernization initiatives such as mainframe transformation. As enterprises look to update legacy systems and integrate them with modern technologies, agents can automate data migration, identify what can be done to modernize existing application code and support compatibility with new platforms. In doing so, they help accelerate modernization while reducing the risk of disruption and error.
Overall, the versatility and adaptability of agentic systems make them invaluable tools for a wide range of applications within mainframe environments, transforming the way organizations operate and manage their mission-critical systems.
78%
of respondents stated that they are implementing or plan to implement agentic AI in the next three years to develop, test and manage business applications.
A security-first mentality
As with any large-scale technology adoption, security is paramount, and agentic AI systems must be built with strong safeguards to drive safe and reliable operations. Role-based access control, for instance, assigns permissions to agents based on their functions, thereby reducing the risk of unauthorized actions. These measures are reinforced by secure storage, encryption and multi-factor authentication, all of which help protect system credentials and end-user authorization.
Equally important are safeguards around how agents interact with data and tasks. Input validation, output monitoring and behavioral constraints help prevent misuse of language model components, while human oversight of critical operations adds another layer of protection, ensuring that disruptive or high-stakes actions are carefully reviewed before execution.
Finally, continuous monitoring, observability and auditing provide transparency and accountability, strengthening the overall integrity of agentic systems and helping organizations maintain trust in their deployment.
The road ahead
The adoption of agentic AI systems is expected to grow exponentially, with nearly 80% of respondents to the 2025 State of Mainframe Modernization Survey stating that they are implementing or plan to implement agentic AI in the next three years to develop, test and manage business applications. For mainframe organizations, understanding and embracing these technologies is crucial for maintaining a competitive edge in the rapidly evolving AI landscape.
Agentic AI represents a revolutionary shift in technology, offering unparalleled opportunities for innovation and efficiency. As organizations navigate this transformative era, the adoption of these systems will be instrumental in shaping the future of mainframe technology and beyond.
