Databricks Context Engineer Associate Certification Sample Questions

01. A context engineer at a property insurer reviews a trace from a claims-triage agent:

- Step 3: the agent summarizes a policy document, stating the deductible is $500; the document says $5,000. Only the summary is kept in context.
- Step 7: the agent's scratchpad records deductible = 500 and uses it to estimate the payout.
- Step 12: the agent cites its own step-7 note when drafting the customer letter.

Which context failure mode does this trace show?
a) Context distraction, because the long trace made the agent lean on accumulated history over its own reasoning
b) Context confusion, because the policy document added irrelevant content that the agent then used in the letter
c) Context clash, because two sources in the window give the payout estimate conflicting deductible values
d) Context poisoning, because one wrong fact is reused later as if it were true

02. A SaaS support agent retrieves the 25 most similar memories for every turn. Customers with several products now get answers that cite old tickets about products they are not asking about, while the current ticket's history is still needed for good answers.
Which TWO changes reduce over-retrieval without causing under-retrieval of the current ticket's history?
a) Filter retrieval to the account and the ticket's product, so other products' tickets are left out
b) Raise retrieval to the 50 most similar memories, so the relevant ticket history is always among them
c) Turn off memory retrieval and rely on session history alone, so no old ticket can enter the context
d) Raise the similarity cutoff sharply, so only one or two near-identical memories pass into the prompt
e) Rerank the 25 retrieved memories by relevance to the question, passing only the top several to the prompt

03. A logistics agent waiting for customs clearance polled get_clearance_status 14 times over two hours. Each call returned the identical result, status: PENDING, until the 15th call returned status: CLEARED with a release number. The history is about to be compacted.
Which content is safe to remove without affecting later steps?
a) The 15th result with its release number, since the summary can simply record that clearance was granted and when
b) The original customs filing request, since clearance has now been granted and the filing step is complete
c) The 14 identical PENDING results, since they repeat one another and the final CLEARED result supersedes them
d) The agent's note of the warehouse bay where the shipment is held now, since the filing already names the port of entry

04. An energy utility built a single outage-analysis agent that ingests sensor logs from 40 substations, classifies faults, correlates them with weather data and drafts the incident report. The team chose one agent to avoid hand-off losses. Traces show its context climbing past 90% by the twentieth substation, and its later fault classifications cite readings from the wrong substations. Each substation's classification is independent of the others; only the final correlation needs all results.
Which boundary placement failure is occurring?
a) Boundaries are too fine, so details are lost across hand-offs between the analysis stages
b) Boundaries are too coarse, so one agent accumulates every substation's work
c) The agent's weather retrieval is stale, so readings are matched to the wrong time and substation
d) The agent has a poisoned context, because one wrong reading is reused across every later classification

05. A retail chain runs a pricing agent and a promotions agent against the same weekly catalog. The pricing agent marks down winter coats by 30%; minutes later the promotions agent, working from the catalog snapshot it loaded at the start, adds a 25% coupon to the same coats, pushing them below cost. Both agents' final reports are merged only at the end of the run, and both agents must keep running in parallel for the catalog to be ready by the Monday deadline.
Which context propagation change would most likely prevent this conflict?
a) Merge both agents' final reports through a reconciliation agent, so it can resolve any overlapping discounts afterwards
b) Give the promotions agent a longer context window, so the full catalog and its history fit together in one load
c) Rerun the promotions agent after pricing finishes, so it loads a catalog with every markdown already applied
d) Publish each committed price change to shared state, so the promotions agent checks it before acting

06. A university student-services agent already resolves intent before retrieval: a query-rewrite step turns each incoming message into a search query, which is sent to Databricks AI Search, and the results ground the answer. The rewrite step receives only the latest user message. A student who has spent several turns discussing a financial-aid appeal asks, "When's the deadline?" The agent gives an accurate course-registration deadline.
Which change resolves this mismatch?
a) Pass recent turns from the session into the rewrite step, so the new query names the aid appeal
b) Add a reranker after AI Search, so the most relevant deadline document rises to the top of the results
c) Add a relevance judge after generation, so answers that do not match the question are flagged for review
d) Store the appeal as a memory in the memory store, so later conversations can retrieve it when it comes up

07. A data-platform team runs a long migration agent that moves a warehouse in twelve milestones, each requiring the same column-mapping rules. A constraint-adherence scorer stays near 0.95 through milestone 5. The team compacts history whenever it reaches 150,000 tokens, and after each compaction the score falls again within two milestones; reviewers find the summaries, built partly from earlier summaries, have dropped mapping rules.
Which intervention most directly restores performance?
a) Compact twice as often, so each summary covers less history and loses fewer of the mapping rules
b) Keep a structured file of completed milestones and mapping rules, starting each milestone from it
c) Reload the full transcript of milestone 1 at the start of each milestone, so the rules come from their original source
d) Switch from compaction to a sliding window of the last 50 turns, so recent milestone work is kept verbatim

08. A media company's orchestrating agent dispatches five research sub-agents to gather background for a documentary script. Each sub-agent returns its complete working transcript, including every search, page read and intermediate note. The orchestrator now runs short of context before it can assemble the script.
Which sub-agent output change would most reduce the orchestrator's context load?
a) Return the transcript minus tool calls, so only reasoning text reaches the orchestrator
b) Return only a one-word relevance rating per topic, so the orchestrator spends almost no tokens on each result
c) Return the full transcripts in a fixed order and format, so the orchestrator can read them more predictably across its window
d) Return condensed findings, sources and open questions, so the orchestrator keeps only what it needs

09. An airline's customer agent runs a fixed pipeline: (1) retrieve policy documents with the raw user message, (2) classify the user's intent from the message and the recent conversation turns, (3) generate the answer from the retrieved documents. A passenger asks, "Can I change it?" after discussing a seat. The agent returns an accurate summary of the flight-change fee policy.
Which change to the pipeline should the AI engineer make?
a) Move intent classification ahead of retrieval, so the intent it resolves can shape what the query searches for
b) Add a groundedness judge after step 3, so answers not supported by retrieved documents are blocked
c) Retrieve more documents at step 1, so the seat policy is likely included next to the fee policy
d) Tell the generator at step 3 to cover every reading of the request, so all possible meanings are answered

10. An HR team uses an agent to co-draft a sensitive restructuring announcement with a director over a long, continuous chat. The agent writes key points to an external notes file and, each turn, reloads only that file instead of the conversation. The director complains that the agent reintroduces phrasings she rejected several turns ago and loses the tone she has been steering toward. Each reply depends closely on the exchange just before it.
Which TWO statements correctly identify the replacement strategy and the reason the original strategy is insufficient?
a) Notes are stored outside the model, because reading them back adds retrieval latency to every turn
b) Split the drafting across sub-agents per section, because each section can then be written in a clean context
c) Notes keep checkpoint outcomes, but each turn here depends on the exchange before it, so the flow is lost
d) Keep only the last three turns verbatim, because the director's latest feedback matters most for each reply
e) Replace the notes file with compaction of the conversation, keeping decisions, rejected wording and tone