Spellforge to Experience Workspace: Building a skill-driven agent system for DA/AEM

Most AI demos show you the magic. This talk shows you the plumbing.

We are building a skill-driven orchestration system on top of AEM/DA and spent as much time on governance, observability, and developer experience as we did on the AI parts. The core primitive is a skill, not a prompt. Skills have typed contracts, versioning, and policy enforcement at the boundary. An orchestrator agent coordinates them. MCP-backed tools handle the deterministic operations like page analysis, block discovery, and visual diffing. Sub-agents take care of workflows that benefit from isolation or parallelism. Everything is intimately linked together and overseen by governance tools. But how did we come to this productionable outcome?

I used AI aggressively to prototype code paths and generate demo assets with Firefly. That gave me the chance to generate a fast feedback loop on interaction patterns that became a Summit main stage keynote worthy demo. Now we're closing the loop to go to prod with confidence: typed contracts, test coverage, separation of orchestration from domain logic. The productionization work covers retries, error classification, scoped tool access, secret isolation, and a skill authoring model that internal and customer teams can actually use without stepping on each other.

If you are building agent systems for real content operations and you care about what happens when something breaks at 2am, this talk is for you.