An AI agent platform can maintain healthy pods and available APIs while failing to complete the tasks its users depend on. Long-running workflows introduce failure modes that infrastructure health checks cannot capture: execution state lost during worker restarts, duplicate actions following tool-call retries, and growing queues caused by model-provider throttling. Operating these systems reliably requires an architecture that connects Kubernetes infrastructure to workflow progress, dependency constraints, and task outcomes. This session presents a reference architecture for deploying and scaling a shared agentic platform on Kubernetes. It separates platform controls for tenant policy, execution budgets, and admission from workers responsible for agent execution and tool invocation. The technical discussion examines durable workflow checkpoints, recovery after interruption, idempotency contracts for tools with external side effects, and queue-based autoscaling bounded by model quotas and downstream capacity. Three failure scenarios anchor the architecture: a worker terminating during a multi-step task, a model endpoint returning rate-limit errors, and a tool timing out after successfully changing an external system. Each scenario illustrates the trade-offs between retrying, resuming, reconciling state, and escalating for human review. The session concludes with an observability and validation approach that combines distributed traces, task-level service objectives, and fault-injection tests. Attendees will leave with practical patterns for recovering interrupted workflows, containing failures across tenants, and measuring whether their agent platform completes useful work reliably.
Jothsna Praveena Pendyala is an AI and Platform Engineer, researcher, technical speaker, and Co-Founder of DeepAgentLabs. She specializes in generative AI, agentic systems, machine learning, cloud-native infrastructure, and reliable enterprise AI platforms. Her expertise spans retrieval-augmented generation, agent orchestration, Kubernetes, and AI deployment across Azure and AWS. At DeepAgentLabs, she contributes to an open-source ecosystem focused on building observable, reliable, resilient, and governable AI systems for production. Her work combines LangGraph, Model Context Protocol (MCP), evaluation, observability, and responsible AI practices to support enterprise automation and AI-assisted operations. Her research contributions include work published through NeurIPS, IEEE, and Springer. Through research, technical talks, and open-source contributions, Jothsna connects practical AI engineering with the challenges of designing, deploying, and operating dependable AI systems. Pramod B N is a DevOps Architect, Site Reliability Engineer (SRE), independent researcher, Microsoft MVP, and Alibaba Cloud MVP, specializing in cloud infrastructure, AIOps, observability, Chaos Engineering, and resilient Agentic AI systems. He is the Co-Founder and Architect of DeepAgentLabs, an open-source ecosystem focused on building observable, reliable, resilient, and governable AI systems in production. He also contributes to and leads IEEE technical standards development initiatives and participates in IEEE-USA technology policy initiatives, helping shape operational frameworks for AI-native systems. A multi-cloud architect with 24 cloud certifications, Pramod serves as Chair of ACM Dallas SIGAI and leads two Microsoft technology communities: the Dallas AI Agents User Group and the Global AI Open Source Builders User Group. He actively contributes to open-source innovation, technical research, industry standards, and global engineering communities.
