Agents Don't Browse - AEM in the Age of AI
– Presented by Platinum Sponsor Adobe –
They query, summarise, and act on your content without ever rendering the page — and they're already in your logs, not just your roadmap decks. Meanwhile the platform behind those pages is going agentic too: migration, creation, optimisation, and governance are starting to happen with AI doing the mechanical work directly, not through UI flows.
Two shifts, same system, opposite sides — and most of the conversation only looks at one of them.
This talk is the engineering reality behind that, not a demo of AEM "having AI." We're building systems where agents work directly on migration, creation, optimisation, and governance; where capabilities are exposed as reusable skills instead of trapped in UI flows; and where humans stay in control while agents take the mechanical load. We're also feeding signals from AI-driven discovery back into the platform, so visibility, optimisation, and execution become one loop instead of three disconnected tools.
That forces real architectural change: grounded generation, isolated runtimes, open integration surfaces instead of another walled-off assistant, oversight designed in rather than bolted on.
A view from inside the transition — what works, what's still rough, and which patterns are worth stealing if you want agents to improve experience management without turning the whole thing into a science project.
wolf
Does Adobe plan to allow composition with other agents and/or MCPs and/or skills not from Adobe ecosystem? Most shops cannot be ground the context with data solely stored in Adobe systems.
jack.jin
One of the agentic design principles is that agents should be given minimal privileges for the task at hand. An admin might want to perform read only tasks and skills If we grant cx co worker with a bunch of skills and connectors, with the users access, how do we solve the mono agent having too much access? We’re saying isolate blast radius.
Michael
What makes Coworker special? A lot of large companies have already invested in e.g. Copilot/Microsoft. How do you plan to compete with that?
npeltier
Beside skills, how do you get human decisions into agents context?
Thomas
Providing capabilities is nice, but how can responsibility be provided? Consider editors changing content in a large scale - how can they see what actually changed and take ownership for that, in a way that is as effective as changing it?
cziegeler
This is similar to what you know from coding agents, where it shows you a preview (diff), then you can accept, or correct. When it comes to batch changing, it is similar, you either accept each change or trust it after some changes. To add, everything runs in the right context, with brand support.
Lisa Brown
How do you recommend handling collisions across the Adobe ecosystem with Coworker? If an author is working in AEM and a traffic coordinator is working in Workfront, and Coworker commits a change across both, what gets prioritized?
cziegeler
This depends, the action is started by either one of them, for example the content author. if that author does not have write access in workfront, coworker will not be able to change things there. And vice versa
Tomasz
Instead of observability I try to talk about AI explainability - being able to explain what ai did, and why, in a short, easy to inspect way for humans to maintain control and oversight
cziegeler
yes, agree, way better wording and explanations
wolf
Do you plan to provide a choice of models for Coworker?
cziegeler
Unfortunately, I cant speak about potential roadmaps - what I can say is that we are hearing this request from several customers.