Adobe's Data Engineering Agent Promises Weeks-to-Days Onboarding

Anyone who has staffed an Adobe Experience Platform rollout knows how the first six weeks go: someone's writing mapping spreadsheets, someone else is chasing a Marketo export, and a data engineer is rebuilding schemas at 11pm because loyalty IDs arrived as strings. That's the scene Adobe is now trying to automate away. On April 20, Adobe announced the Data Engineering Agent, a new agentic workflow built on Experience Platform Agent Orchestrator that will "soon" be generally available.

What Happened

As Adobe for Business reported in a post authored by Huong Vu, the Data Engineering Agent is designed to streamline data life cycles for engineers, architects and data consumers such as marketing operations specialists. The pitch is blunt: traditional data onboarding into Experience Platform often takes weeks, sometimes months. Adobe says the agent will reduce that to days.

The agent is not yet launched. Adobe describes it as coming soon, powered by Experience Platform Agent Orchestrator, and driven through the AI Assistant conversational interface. Users interact with natural language prompts inside a human-in-the-loop workflow, meaning the agent proposes and the human approves.

Scope is broad. On the onboarding side, it will pull from sources including Amazon S3, Data Landing Zone and Marketo, validate data quality, approve AI recommended fields for schema creation, configure semantic enrichment, and create dataflows using AI Assistant for ingestion into Experience Platform. On the SQL side, it will generate optimised, schema-aware SQL statements with previews before execution, monitor and troubleshoot SQL jobs without handling the queries UI or switching tools, validate dataset readiness before running jobs, and guide users through automatic remediation when things break.

The downstream beneficiaries are the usual Adobe suspects: Real-Time Customer Data Platform, Customer Journey Analytics and Journey Optimizer. Clean data in, faster activation out. Adobe also says the agent will surface operational insights that visualise lineage, dependencies and relationships, plus in-context product knowledge grounded in Experience League, community forums and public GitHub documentation.

My take: this is Adobe playing catch-up to the agentic data engineering story Snowflake and Databricks have been telling for a year, but aimed squarely at the marketing data plane where Adobe still owns the relationship.

Technical Anatomy

Strip away the marketing vocabulary and the architecture is a classic orchestrator plus tool-use pattern. Agent Orchestrator is the planner. The AI Assistant is the conversational surface. The agent itself calls into Experience Platform's existing primitives: source connectors, schema registry (Experience Data Model), dataflows, and the Query Service. The novelty is that a language model is authoring the artefacts a data engineer would normally hand-craft.

Onboarding works roughly like this. The user points the agent at a file or source such as S3, Data Landing Zone or Marketo. The agent inspects the payload, proposes field mappings to XDM, flags quality issues, suggests semantic enrichment, and builds the dataflow. A human approves before publish. That approval gate is doing a lot of work here. Schema drift in a customer profile dataset can quietly poison a CDP for months before anyone notices a segment is undercounting. Gating publish behind review is the right call.

For SQL, the pattern is similar. The agent reads the target schema, generates a schema-aware statement, shows a preview, then monitors execution. If a job fails or a dataset isn't ready, it surfaces the issue and proposes remediation. That's a meaningful shift from the current workflow, where an analyst writes a query, kicks it off, and finds out something broke only when a Journey Optimizer campaign misfires.

The obvious comparison point is the rest of the analytics stack. Teams running dbt already have testing and lineage as first-class concepts, documented in dbt's docs. Snowflake and Databricks both ship agent-style SQL copilots. What Adobe is betting on is that Experience Platform customers would rather keep their data prep inside the Adobe perimeter than wire up an external transformation layer. That bet is reasonable for mid-market marketing teams. It's a harder sell for enterprises who already standardised on a lakehouse.

The uncomfortable read: the agent only helps if your data is already landing in Adobe-friendly shapes. Garbage in, LLM-assisted garbage out.

Who Gets Burned

Three groups need to pay attention. First, Adobe implementation partners. A significant chunk of their revenue comes from billing weeks of data engineering work for onboarding ecommerce transactions, loyalty activity and customer profiles, the exact datasets Adobe calls out as examples. If weeks compress to days, so do invoices. Partners who priced on effort are going to feel it. Partners who priced on outcomes will be fine.

Second, in-house data engineering teams at Adobe shops. Nobody's job disappears, but the work shifts. Less time writing mapping logic, more time reviewing agent output, governing the model's behaviour, and handling the edge cases the agent fumbles. Teams I've worked with that adopted SQL copilots found the review burden was real: reading generated SQL carefully enough to trust it takes almost as long as writing it, at least for the first quarter. Budget for that.

Third, marketing ops specialists. Adobe is explicit that data consumers, including marketing ops, are target users. That's a meaningful expansion of who touches schemas and dataflows. It's also a governance landmine. When a marketing ops analyst can ingest a new Marketo export through a chat interface, the definition of "approved data source" gets fuzzier. Expect the next 90 days of any Adobe rollout to include an argument about who gets agent access and what the approval policy looks like.

Production incidents I've seen across iGaming and fintech almost always trace back to one of two things: a schema change no one communicated, or a SQL job that silently returned the wrong shape. An agent that previews SQL before execution and validates dataset readiness addresses both, if teams actually read the previews. That is a big if when the interface feels conversational and fast.

Playbook for Data Teams

If you run analytics or customer data infrastructure on Adobe, here is the week-one checklist.

Inventory your current onboarding backlog. List every dataset stuck in the pipeline, the source system, and the engineer-days it's consuming. When the agent lands, you'll want a before-and-after you can actually measure. "Weeks to days" is Adobe's claim. Your claim should be specific to your datasets.

Lock down the human-in-the-loop policy before enabling the agent. Decide who approves schemas, who approves dataflows, and who approves SQL execution against production datasets. Write it down. Agentic tools fail safely only when the approval chain is explicit.

Instrument lineage independently. Adobe says the agent will visualise lineage, dependencies and relationships. Good. Don't rely on it alone. Keep an external record of which datasets feed which Real-Time CDP audiences and Journey Optimizer campaigns. If the agent misconfigures something, you want your own map.

Stress-test SQL previews. Before trusting agent-generated queries in production, run a sprint where engineers deliberately review every preview. Catalogue the failure modes. Schema-aware does not mean semantics-aware, and the difference shows up in revenue reporting.

Renegotiate partner contracts if you're paying by the engineer-week. If onboarding genuinely compresses, your statement of work assumptions are obsolete. Verdict: treat the agent as a productivity multiplier for senior engineers, not a replacement for them, and your rollout will be boring in the best possible way.

Key Takeaways

• Adobe Data Engineering Agent, powered by Experience Platform Agent Orchestrator, will soon automate onboarding, SQL prep, data collection and troubleshooting inside Experience Platform.
• Adobe claims onboarding time drops from weeks (sometimes months) to days for datasets like ecommerce transactions, loyalty activity and customer profiles.
• Supported sources at launch include Amazon S3, Data Landing Zone and Marketo, with downstream delivery to Real-Time CDP, Customer Journey Analytics and Journey Optimizer.
• The agent uses a human-in-the-loop workflow with natural language prompts, SQL previews before execution, and automatic remediation guidance.
• Implementation partners billing by effort, marketing ops governance policies, and contract assumptions all need a review in the next quarter.