AI Agent-Driven Content Migration for Edge Delivery Services

Content migration remains the persistent bottleneck in web delivery projects. It is largely unaffected by site budget, by the quality of the messaging, or by the tooling intended to accelerate the work, because the underlying task still reduces to manual copy and paste. Having worked on both the client and integrator side of these projects, we observe that while the methods for building and maintaining sites continue to modernize, migration timelines remain constrained by manual content handling.

This talk presents an AI agent-driven content migration system built for Adobe Edge Delivery Services. The system accepts a range of inputs, including PDFs, static websites, dynamic React pages, and AI-generated content used for scale testing. It produces a single standardized output: DA Live authored pages composed of EDS-ready blocks that are previewable and publishable. Rather than performing one-shot generation, the system operates as a closed loop. Playwright allows the agent to inspect the pages it has produced, DA Live provides the means to create, update, and publish content, and a decoupled evaluation framework scores quality across structure, accessibility, visual correctness, and content fidelity.

The first half of the talk addresses version 1.0, the system as it currently operates. We describe the input strategy, the orchestration layer built on Make.com, and the two MCP integrations. We then cover the canonical block library, a human-readable memory system that improves across successive runs, the context-window strategy, and the self-testing mechanism by which the agent corrects its own output before human review. This section concludes with a live demonstration of the full loop, from PDF input through DA Live output to evaluation scoring.

The second half addresses version 2.0. Version 2.0 introduces an agent-to-agent (A2A) architecture that decomposes the original monolithic agent into a mesh of specialized, independently addressable agents responsible for generation, migration, and evaluation. Each agent can be coordinated, measured, and improved in isolation, with Make.com serving as the orchestration fabric and human-in-the-loop checkpoints retained where they provide the most value. The evaluation framework underpins this work by determining whether each added capability, including alternative models, memory configurations, and prompt techniques, produces a measurable improvement in results.