Adobe Advances Its Agentic Tech Stack

In Adobe’s communications about AI, the word “model” is becoming less common: the era of agents has arrived.

Several have just moved to general availability. The same goes for the underlying orchestrator (reasoning engine + knowledge base manager).

Interaction via the built-in assistant

The agents in question are accessible across various products of the Adobe Experience Cloud suite. Essentially through the chat interface known as the AI Assistant*, which allows the engine to expose its chains of thought. They activate or not depending on the requests. Their features, in broad strokes:

• Audience management (Audience Agent)
• Journey management (Journey Agent)
• Experiment management (Experimentation Agent)
• Site management (Site Optimization Agent)
• Data insights analysis (Data Insights Agent)
• Product support (Product Support Agent)

Some of these agents feed — or will feed — autonomous and/or complementary applications to Adobe products. For example, the Site Optimization Agent operates in the Site Optimizer module, accessible via the Experience Manager CMS console. The one dedicated to test management is offered in Journey Optimizer and will also underpin a related module (Journey Optimizer Experimentation Accelerator), which can also operate in tandem with Adobe Target.

Without customization or third-party models for the moment

The catalog includes other agents, still in experimental state. Among them:

• Workflow Optimization Agent (in Adobe Workflow)
• Content Production Agent (in GenStudio for Performance Marketing)
• Account Qualification Agent (in Brand Concierge, a conversation personalization module currently in early access)
• Product Advisor Agent (the same)
• LLM Optimization Agent (in LLM Optimizer, a module under development to optimize SEO in AI engines)
• Data Engineering Agent

Customizing agents isn’t yet possible. Adobe plans to deliver an environment for that, along with a registry and an SDK. It also mentions the integration of long-term memory… and the coordination of third-party models — something the orchestrator cannot handle yet.

* The knowledge graph leveraged by the AI Assistant is divided into three sub-graphs: internal operational data, external operational data, and Adobe product documentation. By “operational data,” it should be understood that the system doesn’t work directly on client data — and more broadly on personal data — but on metadata. For example, you can ask how many data sets are available, which journeys have been modified after a given date, or which schemas are not linked to any audience.

For additional reading :

RPA: when agent-centric approaches seize business strategy
From UX to AX: designing interfaces for AI agents
AI agents: a taxonomy of protocols beyond MCP
To reason better, must LLMs do without language?
AgentForce, a disruption first in terminology?