Diamond Sponsor of Agentics Day: MCP + Agents, Solo.io is proud be a Diamond Sponsor of Agentics Day: MCP + Agents. We’re committed to helping teams build secure, scalable, and production-ready AI applications by connecting models with real-world tools, data, and workflows. Join us to explore practical implementations, share best practices, and shape the future of safer, more capable AI together with the MCP community.

You might imagine a robot as a standalone entity, like R2-D2, with all processing onboard. In reality, modern robotics systems often send vast amounts of data to cloud backends for processing and training, involving complex networking and distributed computing.

While edge devices and IoT also combine cloud and local computation, robots face unpredictable connectivity, requiring failover to local compute and smaller models. Determining which computations should run on-device versus in the cloud presents rich research opportunities, and distributing computation adds networking and orchestration complexity. Effective system design must carefully balance the efficiency of local computation with the greater power of cloud-based reasoning.

This session explores how cloud-native, distributed architectures can coordinate edge and cloud computation to enable reliable, intelligent robotics in real-world environments.

In this session, we’ll explore the communications patterns behind core AI use-cases - training, inference, serving, consumption, and agentic interactions. You’ll learn the basics of key concepts and features, like MCP, A2A, token-based rate-limiting, prompt guarding, etc, and where they apply. We’ll then discuss where conventional networking technologies - service mesh, CNIs, gateways, and L7 proxies - fit into this paradigm, and when AI-aware implementations like the Gateway API Inference Extension are required for routing, observing, and securing GenAI workloads. Finally, we’ll end with practical guidelines for taking these AI tools to production and integrating them into your stack.

In this demo-driven session, Lin first shows the "hard way": modifying MCP server code, rebuilding, and redeploying to add authentication. After that, she introduces a much simpler method by offloading authentication to Keycloak through an MCP gateway, which allows you to secure your servers without changing or redeploying them.

As AI inference workloads become the norm in Kubernetes, the Gateway API Inference Extension (GAIE) faced a pivotal architectural question: should request scheduling act as a "consultant" via an external processing filter or become a full-fledged Service in the traffic path?This panel brings together maintainers and practitioners to explore two models for routing AI inference. GAIE uses external processing to make scheduling decisions without proxying traffic - an unconventional but deliberate choice that prioritizes simplicity and performance. An alternative would place the scheduler directly in the data path as a next-hop Service, aligning with Gateway API norms.This discussion isn’t just theoretical - the architecture shapes what GAIE can do and how it integrates with Gateway API. Attendees will leave with a clear decision framework for designing GW API extensions, and a deeper understanding of how architectural choices impact performance, complexity, and ecosystem alignment.

In this demo-driven session, Lin first shows the hard way: modifying MCP server code to add OAuth support, then rebuilding and redeploying. After that, she introduces a simpler method that offloads authentication and authorization through an MCP gateway, which allows you to secure your servers using OAuth, JWT, SPIFFE and other identity systems without changing or redeploying them. Finally, she shows how her AI agent interacts with the MCP server tools securely, with proper authentication and access control.

As AI agents evolve from experiments to autonomous actors in complex systems, they strain traditional authorization models. How can we empower them while enforcing least privilege and governance?

This session introduces agentgateway, a dataplane powered by kgateway that supports the Model Context Protocol (MCP) and integrates with CNCF policy engines like Kyverno. Nina and Luc — maintainers of agentgateway and Kyverno — show how Kyverno policies evaluate MCP traffic to prevent uncontrolled tool access and privilege escalation, ensure namespace and tenant isolation, and tie every Kubernetes action to real user identities. Using Kubernetes, OIDC lookups, RBAC, and business-specific policies, it delivers least-privilege control, auditable access, and compliance visibility.

With agentgateway and Kyverno, Platform and Security teams can prove continuous compliance for AI workloads and securely scale AI agents with Kubernetes-native policies and cloud-native best practices.

‍

At Skyscanner scale, every architecture choice carries cost and ops impact. Running dozens of Kubernetes clusters and >60M req/min, we were stuck with costly, complex east-west NLBs. This talk shows how we rebuilt our network to save millions on AWS while preserving resilience and compliance, and how we balanced cost, performance, and auditability.
What we changed: removed east-west NLBs in favor of pod-to-pod multicluster for east-west traffic; added gateways for controlled ingress/egress while using Spot instances; deployed Istio Ambient Mesh for zero-trust without sidecars; and built OpenTelemetry-based observability.

You’ll leave with migration patterns, cost-modeling tips, and ops lessons for teams running Kubernetes at scale.

Challenges & risks:

• NLB sprawl and the associated
• Implementing Istio Multiclsuter
• Operational drag
• Sidecar overhead
• Coordinated rollout across many clusters
• Maintaining peak-time performance while shifting traffic
  

Industry analysis reveals that expensive GPU cards are often utilized for processing simple queries that could efficiently be handled by CPUs, such as batch processing, non-real-time requests, or small context queries.
The presentation will showcase an intelligent routing mechanism built on a three-layer architecture composed of CNCF projects: Istio and AgentGateway as the secure communication layer, LLM-D as the inference framework, and kagent as the agent orchestrator. This enables real-time analysis of request complexity. Simple queries are routed to Small Language Models (SLMs) with fewer than 8 billion parameters.

Curious how kagent brings agentic AI to the cloud native ecosystem? Interested in contributing to one of the fastest growing CNCF Sandbox projects (incubation pending), featured on the latest CNCF Technology Radar for agentic AI? Want to explore the codebase behind the kagent framework, or peek behind the curtain to see how it all works?

This is your chance. Join the kagent maintainers for a hands-on session diving into the codebase and learn how you can help shape the future of kagent, one PR at a time.

During this session, we will cover the architecture of kagent, how agents and MCP tools work together, how to set up your development environment, how to interact with the community, and how to start contributing your first PR to kagent.

‍

Getting started with kgateway can feel daunting due to its layers of testing, from Gateway API conformance to unit, end-to-end, and performance tests. kgateway is a fully compliant, feature-rich Gateway API implementation. Since joining the CNCF, it has grown rapidly, adding OTEL tracing, inference routing, and MCP authentication, making thorough testing essential.

This session will guide newcomers through running and writing tests that validate kgateway’s functionality and Gateway API behaviors. Using virtualized environments that simulate realistic traffic and services, we’ll show how to run kgateway locally, build tests, measure performance, and validate Gateway API behaviors.

Speakers include past LFX mentors and mentees who can share first-hand experience making their first open source contributions. Participants will gain skills in structured testing and reproducibility applicable across CNCF projects. Bring your laptop, join our community, and start contributing to kgateway!

Modernizing legacy applications for the cloud native world is full of challenges, from technical to strategic. This panel doesn't just list the problems such as technical debt, migration risk, skill gaps, cost constraints, and infrastructure misalignment; we jump straight to the solutions. Panelists will share practical strategies for phased migration, emphasizing how these efforts pave the way for Kubernetes adoption. Our discussion will focus on critical questions like:

• How do you assess whether AI-enabled tools are right for preparing workloads for K8s vs manual refactoring?
• When leveraging AI for refactoring, what validation or human oversight processes are needed?
• How do we design a Cloud Native AI reference architecture that runs modernized apps and AI workloads on Kubernetes, and accelerates innovation and time-to-market for new AI capabilities?

Attendees will leave with a practical roadmap to adopt AI modernization tools that balance innovation with safety and oversight.

‍

As application developers accelerate their output with AI-assisted coding tools, platform engineers face mounting pressure to keep pace. Can nondeterministic AI agents help bridge this gap, or are they a troublesome new source of complexity and unpredictability? What common challenges cause these AI initiatives to stall, and what does a “minimum viable” platform foundation for success actually look like? How do teams evaluate agent effectiveness and manage the cost implications of AI in production? How can platform engineers build trust in nondeterministic systems, and how does the human–agent collaboration model differ from traditional team dynamics?
‍

This vendor-neutral KubeCon panel brings together platform engineering leaders, project creators, and technical experts to share practical insights on the real-world impact of AI agents in platform engineering, offering proven patterns for defining golden metrics, avoiding common pitfalls.

‍

The Model Context Protocol specifies how MCP servers expose tools, data, and workflows to agents. The spec was written in terms of single tenant, desktop based use cases. Enterprises need to move beyond this definition of and begin building “MCP services”: secure, remotely accessible, multi-tenant, governed services that expose sensitive business capabilities to AI agents.

In this talk, I'll highlight three challenges that arise:

Onboarding & Discovery: How do you register, approve and safely expose MCP services while defending against tool poisoning and shadow services?

Authorization & Identity: How much of the MCP Authorization spec can be adopted when most IdPs don’t support the RFCs it assumes? What’s the gap between the spec’s design for public SaaS and the reality of enterprise SSO, policy engines, and workload identity?

Upstream Access & Consent: Once an MCP service needs to call enterprise APIs on behalf of a user, how do we govern delegation and prevent credential misuse?

‍