AI
Innovation Stalls Before Modernization Starts
Innovation stalls when engineers spend most of their time on maintenance. Kenneth Kilty of Microsoft describes the default pattern: applications move from on-premises VMs into containers but stay essentially unchanged. He calls it “lift tinker and shift.” Data stays locked inside those applications. When EU regulations like the Cyber Resilience Act take effect this year, organizations that deferred this work face a scramble. Kilty’s first job was fixing Y2K COBOL. He sees the same pressure building again.
Legacy Security Models Break on Kubernetes
Savitha Raghunathan of Red Hat makes the security gap specific: legacy apps depend on perimeter security, but Kubernetes demands zero trust, encryption at every layer. A single change to a legacy codebase takes over a week just to assess risk before touching it. Compliance compounds the problem. Government mandates now require specific cryptographic algorithms, and legacy dependency trees often cannot support them. Konveyor addresses this with shared, vetted ruleset repositories, pre-built migration scenarios organizations can reuse instead of solving each technology-to-technology transition from scratch.
Why One-Shot AI Modernization Fails
Microsoft built Draft, a template-based tool to generate Kubernetes manifests for legacy apps. It could not anticipate the range of real-world applications and failed. They replaced it with Containerization Assist, powered by an LLM, and immediately saw unexpected outputs: the model started suggesting Amazon base images for containers destined for AKS. The fix was OPA Rego, a policy layer that constrains what the probabilistic model can output. Making AI work at scale means turning something probabilistic into something deterministic. One-shotting a migration is nowhere close to possible.
Structured Workflows for Large, Undocumented Codebases
Microsoft’s P50 for Java projects entering GitHub Copilot is 100,000 to 120,000 lines of code. Open-weight and small models fail at that scale. Kilty’s team now recommends commercial models like Claude Sonnet for large-scale codebase reasoning. The harder problem is context: LLMs are not trained on proprietary enterprise libraries. Raghunathan says Konveyor addresses this with custom rule sets that feed the model relevant context, plus a solution server that stores successful migration patterns and reuses them as hints on future migrations, making each pass more targeted.
Removing Grunt Work at Scale
IDE-first modernization, where a human sits in front of Visual Studio Code, does not scale to hundreds of repositories. Kilty’s team is shifting to asynchronous agents: run an assessment across 100 repos, let agents handle dependency mapping and SBOM generation, then hand developers a concrete task list. Raghunathan adds that Konveyor assigns story points to each migration issue, so teams can plan how many applications to tackle per week alongside feature delivery. The target is removing assessment grunt work before developers engage.
Q&A
Who will replace senior engineers if juniors rely entirely on AI and senior engineers retire? Panelists predicted agentic supervisor layers will automate repetitive review work, but humans remain necessary for business-logic decisions on mission-critical code, and mentoring junior engineers into seniors remains the community’s responsibility. ▶ 20:59
How does AI create room for refactoring when teams are already stretched between maintenance and feature work? Asynchronous agents running assessments across repositories remove dependency-mapping grunt work so developers receive concrete task lists rather than starting from scratch. ▶ 27:02
Key Takeaways
- Legacy “lift tinker and shift” leaves data locked, with compliance exposure rising as EU regulations like CRA take effect
- OPA Rego policy layers constrain AI model outputs, turning probabilistic modernization tools into deterministic ones
- Microsoft’s P50 Java migrations run 100,000-120,000 lines; small models fail, Claude Sonnet-class models needed
- Konveyor’s solution server accumulates successful migration patterns and reuses them as targeted hints on future runs
- Asynchronous agents running assessments across hundreds of repos remove analysis grunt work before developers engage
About the Speaker(s)
Daniel Oh is a Java Champion, CNCF Ambassador and TAG DevEX Co-Chair, Microsoft MVP, and Developer Advocate at IBM MWL, with a background in technical marketing, keynote speaking, and published authorship.
Savitha Raghunathan is a Senior Software Engineer at Red Hat working on application modernization technologies. She leads the Kubernetes sig-security-docs sub-project to create security awareness through documentation, and serves as a Konveyor project maintainer leading community engagement.
Duncan Doyle is Director of Product at Solo.io, managing the CNCF-governed K Gateway ingress gateway for Kubernetes. He previously held product, technical marketing, and consulting positions at Red Hat, with a focus on microservices, APIs, and API management.
Kenneth Kilty is a Principal Technical Program Manager at Microsoft focused on cloud native migration and modernization strategy. With over two decades of experience spanning presales, consulting, and product leadership, he drives enterprise adoption of Azure Kubernetes Service.