Context-Engineer-Associate Exam Guide & Practice Questions

A support agent receives a current request to explain a failed invoice export. Its context also contains several pages of old notes about an unrelated dashboard redesign, including prior decisions and implementation details. The relevant export error appears in a recent tool result, but the agent repeatedly recommends changing dashboard filters and never addresses the export failure. The old notes are accurate, and no instructions conflict.

Which context failure mode best explains the agent behavior?

• AContext poisoning
• BContext distraction
• CContext confusion
• DContext clash

A finance operations team uses an agent to answer questions about invoice approvals. Its context includes the current approved procedure and a cached project note from an earlier policy cycle. The two sources specify different approval thresholds, but neither is marked as untrusted or malicious. In repeated tests with the same invoice details, the agent cites one source and then the other, producing opposite recommendations.

Choose two observations that most strongly indicate context clash.

Select 2 answers, then check them.

• AThe trace contains two policy passages with incompatible approval thresholds, and both are presented as current.
• BThe agent changes its recommendation depending on which of the competing passages it cites.
• CIrrelevant department history is included and consumes a large part of the context.
• DA spreadsheet cell contains an instruction to ignore policy, and the agent follows it.
• EThe agent confuses two vendors with similar names even though the policy is clear.

A company uses a Databricks agent to answer questions about monthly support revenue. Its system prompt says to be concise and helpful. In testing, the agent correctly reports the revenue values returned by its query, but it also states that a campaign caused a decline even when the results contain no campaign information. The team wants a prompt change that directly addresses this failure without adding a large rule set.

Which system prompt revision best addresses the observed failure while adding the least token cost and maintenance burden?

• AAdd an instruction to distinguish facts present in query results from explanations that are not supported by those results, and state when a cause is unavailable.
• BAdd a detailed glossary of campaign terminology and require the agent to consider every possible cause of a revenue change.
• CAdd several examples of revenue questions, including examples that repeat the same unsupported campaign explanation.
• DRequire the agent to ask a clarifying question before answering any question about a change in revenue.

A team is refining a Databricks agent that answers operations questions. Its current few-shot set contains three routine, clearly worded questions with similar answers. In evaluation, the agent mishandles ambiguous requests, sometimes changes the required response structure, and occasionally asserts a fact absent from its supplied context. The token budget allows three more examples, so the team must favor examples that cover meaningful gaps rather than repeat existing patterns.

Choose three. The token budget allows three additional examples. Which examples should the team add?

Select 3 answers, then check them.

• AA second routine question that uses the same wording pattern and produces the same answer structure as an existing example.
• BAn ambiguous request where the agent asks a focused clarification before selecting an interpretation.
• CA request that requires the agent to follow the required response structure, including the specified field names and ordering.
• DAn example where the agent states that a requested fact is unavailable when the supplied context does not contain it.
• EA long example that explains the company domain in detail but does not demonstrate a distinct response behavior.
• FA paraphrase of an existing routine question that uses different terms but follows the same path and response pattern.

A retail analytics team uses a Databricks agent connected to a Genie space for weekly sales questions. The Unity Catalog table has columns named `amt`, `qty`, and `dt`, with no table or column comments. Business users report that answers sometimes use gross sales instead of net sales, and they cannot change the underlying table schema this quarter. The agent has access to a documented view that exposes both measures, but its columns are also undocumented.

Which change is most likely to improve the agent's answer accuracy?

• AAdd precise Unity Catalog descriptions to the relevant tables and columns, including the meaning and units of the sales fields.
• BIncrease the agent's maximum response length so it can provide more detail.
• CAdd the entire sales schema to the Genie space, including tables unrelated to the requested measures.
• DChange the agent's system prompt to tell it to be more accurate when calculating sales.

A finance team is configuring a Genie space for questions about monthly net sales by fiscal quarter. Analysts must follow the company's approved refund and cancellation rules, and fiscal quarters do not match calendar quarters. The catalog also contains several unrelated assets, and the team wants the space to include only objects that directly support reliable answers to this reporting need.

Choose two Unity Catalog assets that should be added to the Genie space.

Select 2 answers, then check them.

• AThe managed orders table, which contains the order date, customer identifier, and net sales amount.
• BThe curated sales view, which applies the approved cancellation and refund rules used for monthly reporting.
• CThe employee performance table, which contains staff reviews and is unrelated to customer sales.
• DThe fiscal-quarter SQL function, which returns the company fiscal quarter for a date.
• EThe inventory snapshot table, which records on-hand quantities but has no sales measures.
• FThe customer support parameterized query, which searches ticket text using a supplied phrase.

A Databricks agent searches an AI Search index built from a Unity Catalog table of internal procedures. Each row has a `body` column with the procedure text and a `department` column with a short label. Users searching for specific troubleshooting steps receive results that share a department name but do not contain the requested steps. The index configuration uses `department` as its embedding source column.

Which remediation most directly addresses the retrieval problem?

• AConfigure the index to generate embeddings from the document body text, then update or rebuild the index so the vectors reflect that text.
• BIncrease the number of search results returned so the agent can inspect more department values.
• CAdd the department field to the search query as a required filter for every request.
• DIncrease the agent's response token limit to allow it to include more retrieved records.
• EReplace the document body with the department field in the source table.

A customer-support agent is packaged as an MLflow model and uses Lakebase for persisted user facts. During a session, a customer approves a rule that replacement shipments must use a staffed pickup location. Weeks later, a new session must apply that rule, but the agent currently retrieves only messages from the active conversation. The support team requires the preference to survive new sessions.

Which memory strategy should the team use?

• AKeep the rule only in the active conversation history.
• BStore the rule as persistent user memory in Lakebase and retrieve it in later sessions.
• CInclude the rule only in the next request prompt.
• DRead the prior session's MLflow trace as the agent memory source.
• EWrite the rule to an in-context scratchpad that is cleared after the response.

A Databricks operations agent, logged as an MLflow model, processes a reimbursement in several stages. It validates receipts, checks policy, obtains manager approval, and then submits payment. Jobs can be retried on another worker after a timeout, and the agent must resume from the last completed stage without repeating an approved payment. The current design keeps progress only in an in-context scratchpad.

Choose two. Which requirements justify using a Delta-backed state object instead of an in-context scratchpad?

Select 2 answers, then check them.

• AProgress must survive a worker restart and be available when a retry resumes the task.
• BState should exist only during the current model invocation and then be discarded.
• CThe task needs structured, durable updates to record completed stages for later resumption.
• DThe scratchpad must provide durable state across agent invocations without any external storage.
• EKeeping the full conversation in the prompt must guarantee recovery after a restart.

A travel agent is deployed as an MLflow model and reads long-term preferences stored in Lakebase. Each memory row has a user ID, a category, an expiration date, and a short text value. A request asks for the user's active preferred currency, and the application must exclude expired rows and return the exact category match. There is no need to find conceptually similar statements.

Which retrieval approach best fits this request?

• ARun a structured query against Lakebase that filters by user, category, and expiration date.
• BUse Databricks AI Search to retrieve the nearest semantic matches from the memory text.
• CScan the current conversation history for a similar currency preference.
• DUse an in-context scratchpad to infer the active preference.
• ERead the MLflow trace for the user's prior sessions.

A support team connects an agent to an MCP server with 28 tools. The agent uses only three tools for a particular ticket workflow, but the client currently inserts every full tool description and input schema into each request. The context window is shared with a long ticket transcript, and the team can change how tool metadata is exposed without changing the tools themselves.

Which approach best reduces tool metadata in the agent context while keeping the tools discoverable and usable?

• ALoad every full tool schema at the start of each session and retain it for all requests.
• BExpose only the three tools used in the current workflow, with no way to discover other tools.
• CShow a compact catalog of tool names and brief descriptions first, then load the full description and schema for a selected tool before invocation.
• DOmit the input schemas and let the agent infer the parameters when it calls a tool.
• ECopy every full schema into the system instructions instead of providing it through the MCP client.

A customer operations agent uses two MCP tools. The first, find_customer_by_email, searches for a customer and returns a list of possible matches. The second, get_customer_record, retrieves a record using a customer ID. Their current descriptions both say that they find customer information using identifying details, and the agent often calls the retrieval tool when given an email address.

Choose two changes that would most directly reduce confusion between these tools.

Select 2 answers, then check them.

• ARewrite both descriptions to use the same wording so the agent sees a consistent style.
• BDescribe find_customer as searching by email or name and returning candidate customer IDs, without retrieving a full record.
• CAdd both complete input schemas and all optional parameter details to the agent instructions.
• DDescribe get_customer_record as retrieving one record by exact customer ID, without searching by email or name.
• EGive both tools the same name so the agent can select either implementation.
• FTell the agent to select whichever tool appears first in the MCP tool list.

A team uses a Databricks agent to plan a migration of a Delta table used by a reporting pipeline. The trace records that a full refresh would exceed the nightly processing window, so the team rejected that approach and chose an incremental update. After compaction, the summary retains the table name and the chosen goal but omits the rejection and its reason. The agent later recommends a full refresh.

Which category of information should the compaction process preserve to prevent the agent from repeating the rejected approach?

• AThe user’s general preference for concise answers
• BEvery raw SQL result and tool response from the trace
• CRejected alternatives and the reasons they were ruled out
• DThe complete sequence of intermediate reasoning statements

A Databricks agent reviews a long job trace before recommending a repair for a failed Delta table update. An early log entry contains a schema migration warning, while later tool responses contain repeated status details. The current compaction prompt keeps the final action and selected results, but the agent misses the warning and proposes an incompatible repair. The team needs a prompt revision that captures relevant trace evidence without keeping every repeated output.

Choose two. Which revisions best tune the compaction prompt for this trace?

Select 2 answers, then check them.

• AKeep only the most recent tool response and the final proposed action in each summary.
• BFirst preserve all trace details that could affect the result, including warnings, constraints, decision rationale, and unresolved uncertainty.
• CUse a fixed summary length and discard the oldest trace entries whenever the limit is reached.
• DRemove all tool outputs during compaction because the agent can rely on its conclusions alone.
• ECheck the summary against the trace for missed relevant evidence, then remove duplicate payloads and unrelated outputs while retaining support for the retained facts.
• FCopy the complete trace into every summary so no evidence can be lost.

An engineering group runs a Databricks agent through several stages of investigating a failed Delta table update. The trace includes user constraints, job status responses, and an older schema warning that becomes relevant only when the agent selects a repair. A fixed rule keeps the latest twelve entries and removes earlier entries when the context window fills. The team wants to know whether this rule is adequate for the task.

Which assessment of the fixed trimming rule is best?

• AIt is sufficient because the most recent entries always contain the information needed for the next action.
• BIt is insufficient because relevance depends on the task, so the compactor must retain evidence linked to requirements, decisions, and unresolved issues rather than trimming only by position.
• CIt is sufficient if the rule keeps all tool responses and removes all user messages.
• DIt is insufficient only because the context window should be increased until no entries need removal.

A Databricks analytics team uses a coordinator agent to prepare a quarterly performance report. It assigns separate analysis tasks to finance and operations specialist agents. Both specialists return plausible findings, but finance uses the reporting period specified in the original request while operations applies a different period mentioned earlier in the coordinator's work. The dispatch logs show that each specialist received only a short task message.

What is the most likely cause of the conflicting specialist outputs?

• AThe specialist agents have context windows that are too small for their assigned tasks.
• BThe coordinator dispatches each specialist only its individual task message and omits the accumulated agent trace.
• CThe specialist agents have different tool permissions.
• DThe agents use different sampling settings.

A product analytics group uses a Databricks coordinator agent to prepare a launch recommendation through several specialist agents. The first specialist establishes that the launch must comply with a regional pricing constraint, but the next specialists receive only their task descriptions and produce recommendations that violate it. The team wants to improve consistency without increasing the agents' context windows or changing their tools.

Which two changes to the dispatch context are most likely to reduce conflicting recommendations? Choose two.

Select 2 answers, then check them.

• AInclude the original request constraints and relevant prior findings in each specialist's dispatch context.
• BIncrease each specialist's context window while keeping the dispatch messages unchanged.
• CInclude the coordinator's relevant prior decisions alongside each specialist's assigned task.
• DAsk each specialist to interpret the request independently so its reasoning is not influenced by other agents.
• ERemove earlier decisions from the dispatch context so specialists focus only on their assigned tasks.