Databricks Context Engineer Associate

The exam

45 scored questions · 90 min · pass mark not published · July 29, 2026 (GA) guide. Mostly scenarios: pick the fix for the cause, at the least token and maintenance cost.

Weights

Names

Vector Search → Databricks AI Search · Genie spaces → Genie Agents · AI Gateway → Unity Gateway · Knowledge Assistant, Supervisor Agent: legacy (replaced by Genie Agents).

Failure modes

Rules

• Attention budget: cut what costs every turn and matters on few (rare rules, unused tool schemas, consumed raw output).
• Degradation starts before the hard limit: segment MLflow eval scores by context length, intervene before the drop.
• Prevent first: minimal tool set · just-in-time retrieval · tool result scoping (parameters for fields and rows).
• Stack: governance, metadata → UC · per-user state → Lakebase · tool protocol → MCP · see and score context → MLflow 3.
• Reasoning: extended for interdependent multi-step work; reduced for routing or classification on tight latency or budget. Reasoning tokens add cost and latency; Controls: hybrid Claude budget_tokens < max_tokens; adaptive Claude an effort level; some others reasoning_effort.

Genie: symptom → asset

Test with benchmarks; Genie doesn't use them as context.

Prompts

• Few-shot: keep by marginal contribution (untested tool path, output structure, ambiguous input); drop duplicates.
• Over-refusal from a topic ban → forbid the specific behaviour.
• Patch pile-up → consolidate into one rule, delete patches.
• Bigger prompt justified when the gain clearly beats run-to-run variance on the same dataset and scorers; ablate to find the element that drives it.

Metadata and curation

• Genie reads UC column names and descriptions: fix coded values with column comments, near-twin tables with table comments. COMMENT ON COLUMN needs DBSQL or Runtime 16.1+.
• Trusted: tag system.certification_status = certified / deprecated.
• PII: data classification adds class.* tags; ABAC policies on governed tags mask or filter.
• Genie space: one domain, gold views with described columns, trusted queries; up to 50 tables, views or metric views. Knowledge store edits don't change UC.

AI Search

Retrieval: symptom → fix

Strategy

• Chunking: headings → structure-aware · code and tables → element-aware · prose → fixed or semantic.
• Pre-inference for stable docs, single turn · just-in-time tool (UC function, SQL) for live or exact values · both for mixed.
• RAG: UC volume → chunked Delta table → Delta Sync index → inject top chunks with citations.
• Authoritative only: lineage → tag the certified set → dedicated service principal with grants on it alone.

Memory

• Lakebase: managed Postgres (OLTP). Agent sessions (Beta): one interaction, keyed by actor_id (required) and session_id. Agent memory (Beta): long-term, BM25 + semantic search. Self-managed LangGraph checkpointing: legacy. MemorySaver is in-process; Lakebase-backed state survives restarts.
• Durable state object over a scratchpad: spans restarts or days, outgrows the window, shared.
• Persist distilled facts, decisions, open questions; not transcripts.
• Static for small stable facts at session start; dynamic for changing or large records.
• Over-retrieval → threshold, cap, drop superseded · under → lower threshold, raise cap, hybrid.
• Accurate but off-target → resolve intent before retrieval.

MLflow 3

mlflow.genai.evaluate(data=, predict_fn=, scorers=) · scorers RetrievalGroundedness, RelevanceToQuery, Correctness (needs expectations). Compare on one dataset version, same scorers; reliable = consistent across runs.

Rules

• Managed MCP (workspace endpoints, Public Preview): Genie, AI Search, UC functions, SQL; resource permissions apply. Unity Gateway MCPs system.ai.* are UC securables.
• Progressive disclosure: names + one-liners → full schemas for a shortlist → call, condense. Most cost scales with tools used; only the one-line index grows with tools registered.
• Confused tools → descriptions share verb and entities; rewrite each: what it returns, when to call it.
• Pick a UC tool on function, inputs, outputs.
• Clear raw output when its content is captured and no pending step needs it, or it's duplicated.
• Agent skills (SKILL.md): name + description in context, loaded on match. For long, self-contained, rare procedures. Unity Gateway Skills: Beta.

Rules

• Trimming (drop oldest, cap outputs) when relevance decays with age; compaction (model summary) when early detail stays load-bearing.
• Lost after compaction: contradicts a commitment → decisions · forgets work → open items · repeats steps → task state (checklist) · breaks a caveat → constraints.
• Safe to remove: consumed raw outputs, duplicates, superseded drafts.
• Tune: recall first, then precision.
• Aggressive vs conservative per workload, from evaluation; category-aware does both.

Rules

• Inconsistent or conflicting outputs → agents lack shared context; check each agent's trace.
• Sub-agents ignore planning decisions → dispatched with a bare task; attach a compressed decision summary (not the full trace), or a shared Lakebase record.
• Orchestrator saturation → sub-agents return structured summaries (conclusions, confidence, refs); full artifacts to Delta.
• Boundaries too fine → handoff compression overhead · too coarse → one agent's window keeps growing.
• Long horizon: dependent steps → one agent, compaction + notes · independent items → isolated sub-agents · exact figures over phases → structured notes in Lakebase.