Databricks、単一プロンプトから効果的な Genie エージェントを設計する手法を発表

Ask a generic agent about revenue, and it’ll likely grab the first revenue table it finds, not the one finance actually maintains. Your first instinct might be to fix the prompt. But the real problem often isn’t the prompt itself; it’s that the agent is missing the business context needed to interpret the request.

Databricks Genie Agents solve that context problem by allowing you to curate domain-specific agents that can reason over structured data as well as documents and files to automate work. A Sales Opportunity Data agent might surface pipeline risk from governed CRM tables, a Logistics Management agent could track shipments and flag delays across supply-chain data, and a Product Line Lookup agent could answer detailed product questions from an internal catalog. Previously, creating Genie Agents required manual configuration. Now, effective Genie Agents can be spun up using Genie One or Genie Code with a single prompt, leveraging context from Unity Catalog along with the user’s conversation.

How a simple prompt creates an accurate starting point

You can create your first agent in minutes, starting with a single prompt. In many cases, the prompt only needs to describe the desired outcome of the agent, and point Genie to the relevant source of information.

Consider this example: “Use our incident runbooks and service-health data to create an agent that helps support engineers investigate production incidents.”

The prompt is simple, but it gives Genie enough information to get started. It identifies the problem and points to the sources the agent should use.

Of course, the quality of an agent depends on the quality of its context. Genie Agents reason over the data, documents, and files you have already governed in Unity Catalog, so the real work is curating those trusted sources. Once they are in place, Genie can assemble them into an agent for you. For an IT agent, that context might include FAQs, support documentation, incident runbooks, system inventories, and service data. For a data agent, it might include governed tables, metric definitions, dashboards, data-quality rules, and internal documentation.

Context isn't limited to structured data. Genie Agents can also reason over unstructured data stored in Unity Catalog volumes, such as PDFs, Word documents, presentations, and images. When someone asks a question, the agent retrieves the most relevant file content and reasons over it together with your governed tables to produce an answer, always using the permissions of the person asking. That means an Incident Investigation agent can draw on the actual runbook PDFs, and a Product Line Lookup agent can pull from the real product documentation.

Start narrow and expand

Although Genie Agents can support complex, multi-step workflows, the best starting point is usually a focused, recognizable problem that requires repetitive analysis or work. A narrow use case makes it easier to see if the agent is grounded in the right sources and behaves as expected, while making scoping and benchmarking more meaningful.

Take the Incident Investigation agent from earlier. Before rolling it out, run it against a handful of past incidents where you already know the root cause, and check not just whether it reaches the right conclusion, but whether it cites the correct runbooks and service-health data to get there. If it leans on an outdated runbook or the wrong service, that points to context you need to curate, not a prompt you need to rewrite.

Genie Agents include built-in benchmarks for this. You can define a set of test questions, each with an expected answer, and run them to get an accuracy score for the agent. Re-running that benchmark after each change turns a vague sense that the agent “seems better” into a measurable number you can track, and the monitor tab surfaces the real questions and feedback users submit, so you can fold new gaps back into the benchmark as you expand.

Once your first use case is operational, you can add more knowledge and data sources, with each expansion building on a version that has already been tested. The Sales Opportunity Data agent might grow from surfacing pipeline risk to drafting deal summaries or answering questions about win rates by segment. The Logistics Management agent could expand from flagging delays to recommending reroutes or monitoring carrier performance over time. The Product Line Lookup agent could move from answering catalog questions to comparing products or flagging gaps where documentation is missing.

Good starting points include answering questions from a product FAQ, investigating a common type of production incident, explaining cloud-cost changes, or monitoring a defined set of data pipelines. The common thread across use cases and departments is that each agent begins with one recognizable job.

The real asset is the context

Single-prompt agent creation does not replace documentation, or domain expertise. It simply gives teams a faster way to turn those investments into something people can self-serve. The better the foundation, the better the agent. Curated Unity Catalog semantics, such as metric views, domains, and certification, are what the prompt actually draws on to produce consistent, trustworthy answers.

The prompt starts the process, but the quality of the agent still depends on the knowledge and systems behind it. This is where data and IT leaders make the difference: the investment put into a governed, well-defined context is what pays off directly in agent quality.

Build your first Genie Agent

Effective agents start with a clear business need and the trusted context required to address it.

The fastest way to see it is to start with one focused use case. For example:

• "Use our governed CRM tables to create a Sales Opportunity Data Agent that surfaces pipeline risk."
• "Use our supply-chain data to create a Logistics Management Agent that tracks shipments and flags delays."
• "Use our product catalog to create a Product Line Lookup Agent that answers detailed product questions."

Genie Agents make it possible to turn existing data, documents, and workflows into a domain-specific agent from a single prompt, then scope, benchmark, and expand it over time.