Adobe's Agent Architecture Splits Document Understanding from Creation

Adobe launched its productivity agent on May 6 with an architectural choice that caught the attention of enterprise AI teams: complete separation between document understanding and document creation.

The system treats these as two distinct problems. One set of modules parses existing documents, extracting structure and meaning. A separate set generates new content. They communicate through a shared data layer but never overlap in function.

Why Split Understanding from Generation

Most document AI systems bundle parsing and generation together. A single model both reads PDFs and writes new ones. Adobe went the opposite direction.

"We battle inboxes, scroll feeds and navigate hours of video," Adobe noted in their launch announcement. Their solution: specialized agents for each task rather than one agent trying to do everything.

The parsing modules handle unstructured data that Adobe Real-Time CDP now supports, including call center logs and chat transcripts. These modules don't generate anything. They extract.

The generation modules create new documents but don't parse existing ones. They work from the extracted data, not the raw documents.

Three Orchestration Models

Adobe offers three ways to coordinate these split modules, according to Adswerve's analysis of the architecture.

The first model gives agents full autonomy. They decide when to parse, when to generate, and how to sequence operations. This works for simple workflows with clear triggers.

The second model uses human-in-the-loop orchestration. Agents suggest actions but wait for approval. A marketing manager might approve the agent's plan to extract insights from customer calls before generating a campaign brief.

The third model follows predetermined workflows. Agents execute a fixed sequence: parse these documents, extract these fields, generate this output. Less flexible but more predictable.

What Enterprise Teams Are Copying

The split architecture solves a problem many teams discovered the hard way. When one model does both parsing and generation, improvements to one capability often degrade the other. Train it to parse complex tables better, and its writing quality drops. Fine-tune its prose, and it starts missing data during extraction.

Adobe's approach lets teams optimize each module independently. The parsing modules can use computer vision techniques optimized for document structure. The generation modules can use language models optimized for writing quality.

Developer Tech reports that Adobe built this as a platform, not just a product. The CX Enterprise system provides APIs for adding custom parsing or generation modules.

Implementation Details That Matter

The shared data layer between modules uses a schema-based approach. Parsing modules must output structured data matching predefined schemas. Generation modules expect input in these same schemas.

This creates a contract between modules. A parsing module that extracts customer feedback must output specific fields: sentiment score, topic categories, quoted text. A generation module that creates response emails knows exactly what fields it will receive.

Adobe made schemas configurable per workflow. A legal document workflow uses different schemas than a marketing brief workflow. But within each workflow, the schema stays consistent.

Early Results from Adobe's Own Usage

Adobe deployed the agent internally before the public launch. Their marketing teams use it to turn campaign performance data into executive summaries. Their support teams use it to parse customer tickets and generate response drafts.

The split architecture made debugging easier. When generated summaries missed key metrics, they traced the issue to the parsing module, not the generation module. When response drafts used the wrong tone, they knew to check the generation module, not the parser.

What This Means for Your Architecture

Teams building document AI systems face the same choice Adobe faced. Bundle parsing and generation for simplicity, or split them for flexibility.

The split approach requires more upfront design work. You need clear schemas. You need to decide how modules communicate. You need orchestration logic.

But it offers advantages beyond easier debugging. You can swap modules without rebuilding the system. You can use different AI providers for different modules. You can scale parsing and generation independently based on load.

Adobe's productivity agent shows one way to make this split. Your implementation might differ. The key insight remains: treating document understanding and creation as separate problems often leads to better solutions for both.