Agentic Admins: Managing Databricks @ Databricks

The approach aligns with the kind of trusted, self service analytics described in how leading companies deliver AI-powered self-service analytics. For teams wondering where agents fit within a broader platform strategy, the Databricks Platform overview shows how compute, governance, and AI capabilities are unified to support exactly this kind of guardrailed, attention-conserving work.

The operating principle is simple: if a person is clicking, stamping, or typing the same thing for the hundredth time, treat it as a bug, not a job. Conserved attention then funds better questions, better skills, and better paths that compound over time.

Grounding beats cleverness, and data needs it

Early tool calling excelled at combing logs and proposing plausible root causes. Point that same pattern at SQL and failure modes appear: syntactically correct queries that assume column semantics, scan entire tables, and return confident nonsense. Reliability came from ritualized grounding, not clever prompts.

Agents were directed to a specific table, given its documentation, required to check schemas, and to iteratively sample rows to confirm reality matched the docs. For very large sets, a useful tactic was sample, then write negative filters or regex to exclude seen patterns. That let agents traverse billions of rows without flooding context. Data does not explain itself, so the right context must be supplied every time.

Reusable skills and single hat sub-agents

Once a grounding process worked, turning it into a reusable skill delivered consistency. The practical loop was: start from a working example, have the agent write the skill, then run red, green, refactor tests under pressure to expose shortcuts and assumptions. Ask the agent to generate similar test questions to harden the skill further.

Two habits paid off. Bake known good Python into the skill rather than letting the agent regenerate it each run. And give each agent one hat. Writing and reviewing code in one agent made both worse, so responsibilities were split across sub-agents, including a librarian focused on accurate documentation. With skills available in Genie, shared grounding moved into the platform instead of living in one person's head.

Let real questions decide what to build next

Attention saved from no-clicky-ops was spent on questions that matter and are already answerable in system tables. Who actually uses the most critical table, from which surfaces, and how often? Which AI features are catching on by usage, not survey? How quickly do new users become productive? How many downstream people depend on a small number of prolific contributors?

Answers surfaced actionable surprises: agents and chat are fast becoming dominant consumption interfaces, most new users are active within an hour of being provisioned, and a handful of effective question askers quietly shape platform outcomes. Those findings directed investment toward grounding the most used tables for agent consumption and improving onboarding to teach chat-first, question-first workflows.

Build guardrailed paths where agents act and humans govern

Insights translated into "paths," workflows that encode context, permissions, and policy so decisions can be automated or federated. An asset transfer app used workspace identity, Unity Catalog permissions, and org data to auto approve safe ownership changes, route exceptions to the right manager, and remove platform teams from routine loops. Human judgment moved up a level to designing the decision space, not stamping tickets.

To prevent agent created sprawl, the team provisioned scoped sandboxes with clear ownership, schema, workspace folder, service principal, app, and budget. Experiments expired and were reaped if they did not catch on, and cleanly graduated when they did. Three tiers of judgment emerged: auto approve when safe by design, grounded agent judgment for patterned cases, and federated human review as the rare exception.

Start small to unlock compounding attention

The payoff is compounding. Early paths freed attention, agents helped build new paths faster, and cadence accelerated from five paths a quarter to five in two weeks. The invitation is modest and concrete: pick one table, ask one real question, then let the answer dictate the next path to build.