AGNTCon + MCPCon Japan 2026

Modern agent workflows rarely finish in a single tool call. Commerce, travel, healthcare, and enterprise operations all require multi-step, stateful, failure-prone transactions that span agents, tools, and time. MCP today excels at stateless tool invocation, but production systems demand more: resumability, idempotency, compensation, human-in-the-loop, and trust boundaries.
In this talk, we present a Stateful MCP architecture for long-running transactions, using Universal Commerce Protocol (UCP) checkout flows as a real-world case study. We’ll walk through patterns for durable context, agent coordination, transaction lifecycles, rollback strategies, and human approvals, and show how MCP can serve as the orchestration layer for multi-agent workflows that behave more like distributed systems than chatbots.

Agent sandboxes are often described as ephemeral, but real MCP tool workflows are not. Agents edit files, install dependencies, run tests, download inputs, generate reports, and resume after failures. When workspace state is treated as disposable, reproducibility, isolation, and debugging become unnecessarily painful.

This talk presents a practical framework for designing durable volumes in agent sandboxes. Through a structured checklist — covering lifecycle management, tenant isolation, quotas, snapshots, read-only sharing, dependency caches, artifact retention, remount recovery, and audit trails — attendees will learn how to reason about workspace state systematically.

We then map these requirements to Kubernetes PersistentVolumes and explore the tradeoffs using JuiceFS CSI as a concrete open-source example. JuiceFS is used for its strong Kubernetes integration and snapshot capabilities, though the framework applies equally to local containers, serverless environments, and remote tool sandboxes.