AEM 2026.9.0: Canary Deployments, AI Search and Java 25

Adobe shipped Experience Manager as a Cloud Service 2026.9.0 on September 24, 2026, with changes across Assets, Forms and the platform foundation. The ones most likely to change your week are canary deployments, a plan-ahead notice on Java API deprecations and the coming Java 25 runtime, and a batch of AI-assisted features for asset discovery and video. Most of it needs a decision and an owner, not a project.

Key takeaways

• Release 2026.9.0 is dated September 24, 2026. The next feature release, 2026.10.0, is planned for October 29, 2026, so you have about a month to sort out what matters.
• Canary deployments, Java API deprecations and the Java 25 upgrade are the engineering and release-management items. Start there.
• Content Hub adds AI Search, Smart Collections and custom metadata sorting. Teams with big, multilingual libraries benefit most.
• Dynamic Media gets a new video viewer, AI-generated captions and attribute-based access control. Multi-caption and multi-audio support is listed under Limited Availability.
• OpenTelemetry-based APM is Limited Availability and the agent programs are beta. Pilot them, don't depend on them.

What is in AEM 2026.9.0 and who should care?

The notes organise the release around Assets, Forms and Foundation. The themes are AI-assisted authoring and discovery, richer Dynamic Media, and developer productivity. That sounds broad, but the audiences are fairly distinct.

Asset and brand teams get the most visible changes. Content Hub gains AI Search for intent-based, multilingual results, Smart Collections that stay up to date automatically, and custom metadata sorting. Video owners get a new, fully customizable viewer with analytics, AI-generated captions, and attribute-based access control for secure delivery.

Developers and platform owners get a different set. AEM agentic capabilities are now generally available in Adobe CX Enterprise Coworker. The AEM MCP Server adds Cloud Manager operations. IDE agent skills expand with AEM query optimization and more code-migration patterns. On the platform side, there are canary deployments, OpenTelemetry-based APM in Limited Availability, and richer errors for IMS authentication.

The table below maps each change to the people who should look at it first.

How do canary deployments change releases on AEM Cloud Service?

The notes describe canary deployments as a way to validate a release against production before it serves live traffic. That is the headline platform item, and it addresses a familiar worry. A build passes in stage, goes to production, and then meets real content, real traffic and real integrations it never saw.

The notes are brief, so read Adobe's documentation for the mechanics before you commit to a process. What the idea implies is clear enough. You get a chance to see the new version behave against production before customers depend on it. For teams that have been burned by stage environments that don't resemble production, that is a real improvement.

My view: this is worth adopting early if you ship custom code frequently, and worth ignoring for now if you deploy a few times a quarter. The benefit scales with release frequency and with how different your production content is from stage.

What to do:

1. Read the canary documentation and note any prerequisites for your program and pipeline.
2. Decide what you would check during a canary window. Pick three or four concrete signals, such as error rates on key templates, publish flow behaviour and the health of a critical integration.
3. Name who watches those signals and who can stop a rollout.
4. Try it on a low-risk change first, not on your biggest release of the year.

Don't treat a canary as a substitute for regression testing. It narrows the blast radius of a bad release; it doesn't find the bug for you.

What should you do about the Java API deprecations and Java 25?

The notes carry two plan-ahead items: Java API deprecations with a September 28, 2026 date, and an upcoming Java 25 runtime upgrade. Since that date has just passed, the practical step is to read Adobe's deprecation details now and find out whether anything your code calls is affected.

The notes don't spell out which APIs are involved or when the Java 25 upgrade takes effect, and this post won't guess. What you can do without those details is get your house in order.

A sensible inventory looks like this:

• List every custom OSGi bundle, including the ones nobody has touched in two years.
• Note the Java level your build targets and the tooling that compiles and tests it.
• Check third-party libraries embedded in those bundles for their own Java 25 support.
• Run your existing code quality checks and read any deprecation warnings they raise, instead of filtering them out.
• Flag code that depends on an AEM API you know is on a deprecation list.

The new IDE agent skills for code-migration patterns may help with mechanical changes. Use them as an accelerator and review the output the way you'd review a junior developer's pull request. They don't replace a test suite.

This is also the clearest place where an implementation partner earns their fee, if your code base is large or poorly documented. A partner who has done several migrations knows where the surprises hide. If your code base is small and you have tests, your own developers can handle it. See the AEM partner directory if you want to compare who does this kind of work.

How do Content Hub AI Search and Smart Collections help asset teams?

Content Hub now offers AI Search, which returns intent-based, multilingual results. It also gains Smart Collections that stay up to date automatically, and custom metadata sorting options. Together they attack one problem: people can't find approved assets, so they make new ones or grab the wrong version.

Intent-based search matters most where vocabulary varies. A regional marketer searching in their own language for a campaign image doesn't know your internal naming conventions, and AI Search is meant to bridge that gap. Smart Collections matter where your team currently curates lists by hand, such as a launch kit or a set of current logos, and those lists go stale the week after they're built.

None of this fixes weak metadata. Search features sit on top of whatever you've tagged. If half your library has no useful descriptions, better search will surface a messier result set faster. Custom metadata sorting is a small feature but a useful one, because it lets you order results by the fields your business actually uses.

To test it properly:

1. Pick one collection of assets with real volume, such as a single brand or region.
2. Write ten searches that real users have struggled with, in more than one language if you operate in several.
3. Compare results with your current search and count how often the right asset lands in the first screen.
4. Set up one Smart Collection for something that changes often and check it after a couple of weeks.

Don't roll it out to every market at once. Start with the team that complains loudest about search; they'll give you the most honest feedback.

What is new in Dynamic Media video, captions and access control?

There are several related changes here, and they're not all at the same maturity level, so read them carefully.

The highlights describe a new, fully customizable video viewer with actionable analytics, AI-generated video captions, multi-caption and multi-audio track support, and attribute-based access control for secure delivery. In the detailed sections, the multi-caption and multi-audio item for Dynamic Media with OpenAPI capabilities appears under Limited Availability, and so does a native Dynamic Media Template component for Adobe Journey Optimizer emails. Limited Availability means you don't simply switch it on; check eligibility and terms with Adobe.

The practical value is easy to see. Multiple caption and audio tracks let one primary video carry language-specific versions, instead of one asset per locale. That cuts duplication and helps accessibility requirements. AI-generated captions give you a first draft, and a first draft still needs a human check for product names, jargon and anything legally sensitive.

Attribute-based access control matters if some videos shouldn't be public: partner-only training, embargoed launches, licensed footage. Before adopting it, list which videos have restrictions today and how you enforce them, because a control is only as good as the rules you feed it.

For the Journey Optimizer template component, the notes describe marketers adding and personalizing Dynamic Media templates in the email editor, including open-time countdowns. It's Limited Availability, so treat it as a conversation to have with your Adobe contact, not a plan for next month's campaign.

What to ignore for now: the video analytics claims until you've seen the data in your own setup. "Actionable" is a description, not a result.

Where do OpenTelemetry APM, agents and MCP fit in?

OpenTelemetry-based APM is listed as Limited Availability. If your operations team has already standardised on OpenTelemetry elsewhere, ask about access, because a consistent tracing approach across AEM and the systems around it is valuable. If you haven't, it isn't a reason to start.

The agent items are mixed in maturity. AEM agentic capabilities are generally available in Adobe CX Enterprise Coworker, while the notes also point to an Agents in AEM beta program covering production, governance, optimization, discovery and development. The beta warning is blunt: beta releases may contain defects, are provided as-is, and features and APIs can change without notice. Keep production workflows off them.

The AEM MCP Server adding Cloud Manager operations is aimed at developers and platform teams who drive AEM through AI tooling. The new agent skills for query optimization are a good fit for one specific chore: reviewing slow or expensive queries. A sensible pilot is to point the tool at a known problem query, compare its suggestion to what your best engineer would do, and keep the human review step.

The IMS authentication rich errors are the quiet winner for support teams. When sign-in fails, clearer errors shorten the path from a ticket to a cause. Update your support runbook when you see what the new messages look like in your environment.

Teams should govern these tools before adopting them. Decide who is allowed to connect an agent to Cloud Manager, what operations it may perform, and how actions are logged. An agent with deployment access needs the same approval discipline as a person with that access.

How should you plan rollout and measure the results?

You don't choose when 2026.9.0 arrives on Cloud Service, but you do choose what you adopt and when. A simple plan keeps the work proportionate.

For measurement, pick signals that existed before the change so you can compare. For search, track how often users find the right asset on the first attempt and how many support requests ask where something is. For video, track caption corrections needed per video and any accessibility findings. For releases, track post-deploy incidents and time to detect a problem. For query optimization, track the query timings you already collect.

Avoid inventing targets you can't justify. A modest before-and-after comparison on one team is more convincing than a dashboard of vanity numbers.

Don't try to adopt everything in the notes at once. Release notes are a menu, not a to-do list. Teams that chase every feature end up with half-finished pilots and no clear owner for any of them.

When does an implementation partner actually help?

Most of this release doesn't need outside help. Reading the notes, trialling search, and trying a video viewer on one page are things your own team can do.

A partner helps in three situations. First, a large or old code base where Java and deprecation work needs to be scoped and tested. Second, a release process that needs redesigning around canary deployments, particularly if several teams or vendors ship to the same program. Third, governance for agents and MCP access, where security, platform and content owners need to agree on rules.

If you do bring one in, write a tight brief: the Java level you target today, the number of custom bundles, your release cadence and the outcomes you want. Our AEM RFP template gives you a structure for that, and the ranked AEM partner list is a place to start your shortlist. If you want a steer on whether you need a partner at all, the advisor can help you frame the question.

The short version: read the notes, assign the Java inventory now, pilot the content and video features where your users feel pain, and treat Limited Availability and beta items as things to explore, not to depend on. Confirm details against Adobe's own release notes and documentation before you commit budget to any of them.

Sources

• Adobe Experience League: Current Release Notes for Adobe Experience Manager as a Cloud Service — Release notes for the 2026.9.0 feature release, last updated September 24, 2026.