Sign in Subscribe
AI in Marketo

🎧 Developing on Adobe Agent Orchestrator: What the Technical Landscape Looks Like Today

A deep technical read on what Adobe Agent Orchestrator actually is, how it connects to AEP and Marketo, and what practitioners need to understand before any of this lands in their stack.

Juan Pablo Garcia

• 7 min read

I've dropped my deep research document into Notebook LM to produce an audio file to learn at my own pace in audio format. The following post is an AI generated blog post using that deep research document as source. I think it is a good starting point if you agree with my idea: the Agentic AI future will happen inside secure and stable platforms, not by Indie Developers killing the SaaS industry.

So this is an introduction to learn what Agent Orchestrator is, and how the future development by external contributors could look like. I'm sure there will be big announcements at Adobe Summit 2026, most concepts here will remain the same, and some could change. Stay tuned!

audio-thumbnail
Adobe Agent Orchestrator and Job Architecture
0:00
/2834.924263

🎧 This post has an audio companion. Listen to the extensive summary for a full 47 min explanation.

Adobe Agent Orchestrator is not a chatbot upgrade. It is the layer that will eventually govern how Marketo Engage and the other AEP components participates in multi-agent workflows — the coordination surface sitting above your execution tools, routing intent across the products you already license. Most practitioners won't touch this directly for another 12–18 months. But the architecture is being locked in now, and Adobe Summit 2026 is the moment Adobe will formalise a lot of what is currently in early access or announced-but-not-shipped. Understanding the technical landscape before that week matters.

This post is a close read of what Agent Orchestrator actually is, how it connects to AEP and Marketo, and what practitioners in a marketing ops role need to understand before this lands in their stack.

What Agent Orchestrator Actually Is

Adobe Agent Orchestrator is the agentic layer inside Adobe Experience Platform. Its job is to coordinate four components: a conversational interface (what users interact with via AI Assistant), a reasoning engine (the planning brain), a set of specialist functional agents, and a knowledge base that gives those agents enterprise context. When you type a natural-language prompt into AI Assistant today, Agent Orchestrator is what converts that prompt into a structured plan and routes execution across the relevant Experience Cloud applications you have licensed.

The unit of work the platform tracks is an agent job — a series of tasks and actions an agent executes to achieve an outcome. Jobs consume AI credits tied to your existing Experience Cloud licences. This is the economic model: a single-step answer costs fewer credits than a multi-step, multi-agent plan. The credit model also signals the governance intent — every job executes inside the permission model of the underlying products.

The current generally available catalogue is made up of Adobe-provided agents surfaced in Experience Cloud applications: Audience Agent, Data Insights Agent, Journey Agent, Product Support Agent, and a handful of others. These are the agents most practitioners will see first. The more developer-facing tooling — Agent SDK, Agent Registry, and Agent Composer — has been announced and is emerging, but sits in a different maturity tier. Partner access is being staged through an Explorer Program.

Where Marketo Fits in This Stack

The honest framing: Agent Orchestrator coordinates above Marketo Engage. Marketo is a downstream execution surface. It receives instructions; it does not orchestrate them. If an agent job involves sending an email programme, triggering a Marketo campaign, or updating a lead record, Marketo is where that action lands — but the reasoning and routing happen at the orchestration layer.

The adjacent product practitioners should watch is AJO B2B (Adobe Journey Optimizer B2B Edition). AJO B2B handles real-time multi-channel orchestration and is more tightly coupled to the agentic layer than Marketo is today. The plan-based reasoning engine can route actions across products including Marketo, but the integration depth varies by product.

The mental model that matters: think of Agent Orchestrator as a coordination layer that interprets goals and delegates work to execution tools. Marketo is one of those tools. It will not be replaced by this architecture — it will be instrumented by it.

The Six Pillars You Need to Understand

Adobe's architecture documentation describes six pillars for the agentic platform. Here is what each means in practitioner terms:

  • Conversational UX. The interface layer — AI Assistant today, purpose-built agent UIs in future. This is where practitioners will interact with agent jobs, review plans, and approve or reject actions before they execute.
  • Reasoning engine. The planning brain. It takes a goal from a prompt, generates a structured plan (goal → phases → steps → constraints/validation criteria), and can adjust that plan as conditions change. Critically, the plan is exposed in the UI — users can intervene before steps execute. This is not a black box.
  • Functional agents. Domain-specific agents that execute work within their scope. Domain languages like SQL are used internally. Multiple instances of an agent type can coexist. AI Assistant shows the SQL it used to answer a query — that transparency is a design principle, not an accident.
  • Knowledge base and services. Enterprise context — Adobe product documentation, customer metadata about business objects, analytics data. Tenant isolation means your data is not mixed with other customers' data and is not used to train foundation models.
  • Extensible platform. The layer that will eventually allow partners and customers to register custom agents, configure their behaviour via Agent Composer, and expose new actions. The tooling here (Agent SDK, Agent Registry) is what is emerging rather than GA.
  • Trust and responsible AI. Governance baked into the architecture — not bolted on. Agents respect product-level access controls, ABAC policies, and entitlements. The system checks permissions before executing actions. More on this below.

The reasoning engine pillar deserves extra attention. The explicit plan structure (goal → phases → steps → constraints) is not incidental — it is how Adobe is making agentic behaviour auditable. When an agent generates a plan, that structure is visible to the user before execution. This is the intervention point practitioners will need to understand operationally.

What Is Available Now vs. What Is Coming

The honest answer to "what can I use today" breaks down into two tiers:

Generally available: Adobe-provided agents surfaced in Experience Cloud applications. These include Audience Agent, Data Insights Agent, Journey Agent, and Product Support Agent. If you have the relevant Experience Cloud licences, these are accessible via AI Assistant now.

In early access or emerging:

  • Agent SDK — tooling for building custom agents. Announced at GA milestone (September 2025), deeper engineering documentation still developing.
  • Agent Registry — the catalogue for registering and discovering agents. Companion to the SDK.
  • Agent Composer — a single interface for configuring and managing agents. The lifecycle management surface.
  • Agent2Agent (A2A) protocol — multi-agent collaboration, including with non-Adobe agents. This is the interoperability layer that matters for complex MOps stacks.
  • Explorer Program — partner and developer access to preview capabilities before general availability.

The practical implication: if your team is evaluating whether to invest in Agent Orchestrator development work now, the answer depends on whether you have Explorer Program access. Without it, the preparation work is architecture study and toolchain readiness — not active agent building.

The Trust and Governance Model Matters

For marketing ops practitioners in regulated industries, this section is not optional reading.

Agents in this architecture respect product-level access controls and ABAC policies (Attribute-Based Access Control). There is a documented propagation delay of up to 24 hours for policy changes to take effect across agent jobs — something to factor into any incident response planning. The platform uses AES-256 encryption at rest and TLS 1.2+ in transit. Tenant isolation is a core design principle: your customer data does not leave your tenant context and is not used to train Adobe's foundation models.

For organisations running on Experience Platform Healthcare Shield, the platform is HIPAA-ready. If your Marketo instance operates in a healthcare or similarly regulated context, the Healthcare Shield entitlement determines what agent capabilities are accessible to you — and you should verify this with your Adobe account team before piloting anything.

The permission-first interface principle means agents are designed to check entitlements before executing actions. This is by design, not a future aspiration. Understanding your current ABAC configuration in AEP is a prerequisite to understanding what agents can and cannot do in your environment.

If You Want to Go Deeper

The learning path for practitioners who want to be genuinely ready for what Agent Orchestrator becomes over the next 12 months:

Key Takeaways

  • Agent Orchestrator is a coordination layer that sits above Marketo Engage and other execution surfaces — it routes intent, delegates actions, and manages plans. Marketo is an execution target, not a replacement candidate.
  • The reasoning engine generates explicit, inspectable plans (goal → phases → steps → constraints). User intervention before execution is a design feature, not a limitation.
  • The GA catalogue today is Adobe-provided agents. Custom developer tooling (Agent SDK, Agent Registry, Agent Composer) is announced and emerging — active development work requires Explorer Program access.
  • The governance model is architecture-level, not config-level: ABAC policies, tenant isolation, permission-first interfaces, and Healthcare Shield entitlements are baked in. Know your current AEP data governance configuration before piloting agents.
  • The preparation work that matters now: AEP API authentication, App Builder/I/O Runtime familiarity, and AI Assistant as a live debugging tool for understanding how the reasoning layer behaves.

This post is based on a research document collected ahead of Adobe Summit 2026. The landscape will evolve at Summit — we'll update this post with any announcements that change the picture.

Research and source material compiled from Adobe Experience League documentation and pre-Summit technical briefings.

Marketo Ops Radar covers the Marketo Engage ecosystem for marketing ops practitioners. Posts reflect analysis and editorial judgment — not Adobe official guidance.