Everyone wants a faster, simpler way to modernize and migrate their IT environments. But enterprise cloud migrations are too large for manual execution, too complex for simple automation, and too consequential for unconstrained AI.

Understandably, many leaders are eager to use AI to accelerate and simplify migrations, and some may believe the strongest agentic architecture simply uses AI everywhere. But the most successful migrations balance deterministic mechanisms for repeatable processes while reserving agentic AI and human judgment for steps that require reasoning. Most importantly, these processes must be governed by controls that prevent automation tools and AI agents from introducing critical errors into the new environment.

The Enterprise Migration Challenge

Often, large-scale enterprise migrations involve thousands of virtual machines and millions of lines of code. And many of these projects must be completed on tight deadlines tied to the shuttering of a datacenter or the sunsetting of a contract with a key vendor.

Complexity is another key challenge. Mainframe modernization, for example, may involve proprietary code and intellectual property restrictions that require a clean-room approach, where developers must recreate the software functionality without accessing the original source code. Some applications may rely on technologies that cannot easily move to the cloud, forcing teams to decide whether to modify their existing applications or obtain the same functionality in another way.

These decisions require genuine judgment, not just execution, and the consequences of a misstep can be high. Production failures involving critical insurance, telecommunications, or financial workloads can create significant problems and cost substantial amounts of money to remediate.

Agentic AI offers organizations a way to address these challenges. But again, “AI everywhere” isn’t the answer. Instead, migration teams should use agentic AI in parts of the process where the answers aren’t clear from the outset. For example, AI agents can analyze application dependencies, determine which workloads belong in the same migration wave, and right-size cloud resources. Meanwhile, repeatable tasks such as dependency graph calculations, data row counts, and CI/CD pipeline execution can be handled through deterministic automation.

The Cognizant Agentic Migration Platform

The Cognizant Agentic Migration Platform (CAMP) is a production-grade, multi-agent AI platform that utilizes MCP to orchestrate a pipeline of specialized AI agents. These agents do more than just execute migration work. They also engage in ReAct (short for “reasoning and acting”) processes where they evaluate information, make a decision, take an action, and assess the result.

So far, CAMP has been mostly used for two main types of migration. The first is migrating mainframe environments to cloud-native AWS environments. The second is migrating VMware workloads to cloud-native workloads on AWS.

CAMP uses four distinct stages to drive a migration project:

• Discovery and Assessment: This is where we discover and assess the source environment that we’re looking to migrate to AWS. The stage involves identifying dependencies, evaluating risks, and flagging any non-portable technologies that may require special attention. The platform uses tools including AWS Transform to facilitate the process.
• IaC and Migration Wave Planning: Using the information from the discovery process, CAMP organizes workloads into migration waves. The platform also generates infrastructure-as-code (IaC) templates using technologies such as Terraform and AWS CloudFormation, then scans the code to ensure it meets the customer’s security requirements.
• Agentic Execution: Specialized AI agents execute migration tasks, including moving workloads and data to AWS and converting legacy COBOL code to Java. When issues arise, agents can automatically roll back and retry a migration up to three times before escalating the problem to human teams.
• AI Testing and Validation: CAMP tests and validates the migrated environments using parallel AI agents, monitors test pass rates, and produces a recommendation report indicating whether the environment is ready for production.

By letting organizations run migration tasks in parallel with AI agents, CAMP can speed up migration, reduce the need for large project teams, and lower overall costs. Controlling token consumption is critical, especially when organizations use large language models (LLMs) to convert millions of lines of legacy code. CAMP addresses this challenge by selecting AI models based on the complexity of each task (rather than automatically defaulting to the most advanced model), by replacing LLM calls with purpose-built tools where appropriate, and by monitoring token usage against established thresholds.

This approach makes migration costs not only more manageable, but also more predictable. For one customer undertaking a 27-month mainframe migration, for example, Cognizant estimated a token budget of $2.6 million, and we agreed to cover any costs that exceed that budget by more than 10%.

Learn More

Recently, I connected with Techstrong Group’s Jon Swartz for a webinar discussion, now available on demand, on how organizations can adopt a production-oriented architecture that combines MCP-based tool integration, selective ReAct reasoning, deterministic automation, hard quality gates, human oversight, and token-cost discipline.

Key takeaways from the conversation included:

• Why agentic transformation is a systems-engineering problem, rather than a feature of LLMs.
• How organizations can build an architecture that allows them to reason selectively, execute deterministically, and enforce control technically.
• What determines whether agentic AI can scale, including discipline around quality, token usage, and integration.

Watch the on-demand webinar here to learn more.