DCEA

Free DCEA practice questions

10 questions from our DCEA practice tests, each with the correct answer and a written explanation. Read them with the answers shown, or switch to quiz mode and answer first.

Back to the DCEA practice exams

Question 1 of 10, Foundations of Context Engineering

In a long debugging session, an agent increasingly repeats actions it already took earlier in the trace (re-running the same queries and re-reading the same files) instead of trying new approaches. The window is very large but not full. Which failure mode and technique BEST match?

  1. AContext poisoning; restart the session from scratch and discard all findings collected so far
  2. BContext confusion; register more tools so the agent has new options
  3. CContext distraction; compact the history so the agent works from a focused summary Correct answer
  4. DContext clash; add an instruction telling the agent never to run the same query twice in a session

Answer and explanation: When accumulated history becomes so long that the model leans on repeating its past actions rather than synthesising new plans, that is context distraction, and compaction to a focused summary addresses it. A no-repeat instruction treats the symptom and adds a rule the model may not follow under distraction. Restarting discards valid findings, and adding tools increases confusion rather than reducing history length.

Question 2 of 10, Foundations of Context Engineering

A support agent is registered with 45 Unity Catalog function tools, but most conversations use only four of them. MLflow traces show the agent sometimes calls a similar-sounding but wrong tool, and the tool schemas consume a large share of every request. Which proactive change MOST directly reduces context window pressure and tool confusion?

  1. AGive the agent a minimal task-scoped tool set and load other tools only when a request needs them Correct answer
  2. BEnable compaction after every turn so the tool schemas are summarised into a shorter description of the available tools
  3. CSwitch to a model with a larger context window so all 45 schemas fit
  4. DAdd a paragraph to the system prompt reminding the model to read every tool description carefully before choosing one

Answer and explanation: Keeping the tool set minimal and loading other tools just in time removes both the token cost and the ambiguity that causes wrong tool selection. A larger window leaves all the overlapping schemas in place, so attention is still diluted. A reminder in the prompt adds tokens without removing ambiguity, and compaction targets message history, not tool definitions that must be exact to be callable.

Question 3 of 10, System Prompt and Instruction Design

An agent's system prompt injects the current date and the user's region at the very top, followed by 4,000 tokens of static instructions. The team wants to benefit from prompt caching. What is the BEST change?

  1. APlace static instructions first and append dynamic values after them Correct answer
  2. BRemove the date and region so the prompt never changes between calls
  3. CSplit the static instructions across several user messages instead
  4. DInject the date and region in both the beginning and the end of the prompt

Answer and explanation: Prompt caching reuses a stable prefix, so the static content should come first and per-request values later; changing values at the top invalidate the cached prefix every call. Removing the date and region loses needed context. Duplicating them still breaks the prefix, and moving instructions into user messages weakens their role without improving cacheability.

Question 4 of 10, Knowledge Retrieval and Genie Configuration

MLflow evaluation of a RAG agent shows high retrieval relevance scores (retrieved chunks are on-topic) but low retrieval sufficiency: answers miss required facts. Inspection shows key facts often appear in the chunk immediately after the one retrieved. Which remediation MOST directly addresses this?

  1. ALower the similarity threshold so more, less relevant chunks are returned
  2. BRe-chunk the corpus with overlap or larger, section-aligned chunks so related facts stay together Correct answer
  3. CAdd a reranker on top of the current results so the most relevant of the retrieved chunks is placed first
  4. DSwitch to a different embedding model, because high relevance with low sufficiency means the vectors are inaccurate

Answer and explanation: Relevant chunks that lack the needed facts, with those facts sitting in adjacent chunks, point to chunk boundaries splitting related content; overlap or section-aligned chunks fix that. The embeddings are already finding on-topic chunks, so a new model targets the wrong layer. A reranker only reorders what was retrieved, and a lower threshold adds noise without guaranteeing the adjacent chunk.

Question 5 of 10, Knowledge Retrieval and Genie Configuration

A Unity Catalog schema contains curated gold tables and many derived scratch tables created by analysts, some with stale or partial data. A new retrieval agent will query this schema with on-behalf-of-user credentials, and testers notice it sometimes answers from the scratch tables. Which governance strategy BEST constrains the agent to authoritative sources before deployment?

  1. ADelete the scratch tables weekly with a scheduled job so that the agent is much less likely to encounter any stale data
  2. BExpose only curated gold assets to the agent through a dedicated governed schema and tag them as authoritative Correct answer
  3. CInstruct the agent in its system prompt to prefer tables whose names start with gold_ and to ignore all other tables it finds
  4. DRaise the agent's retrieval top-k so that authoritative results outnumber derived ones in every response

Answer and explanation: Constraining the retrieval space through governance (only granting or exposing the curated assets, and tagging them) enforces the boundary regardless of what the model decides. A naming convention in the prompt is advisory and can be ignored or bypassed. Weekly deletion disrupts analysts and leaves a window of exposure, and a larger top-k only adds more context, including more derived data.

Question 6 of 10, Memory Architecture with Lakebase and MLflow

A data-reconciliation agent processes 40,000 invoice records across hundreds of steps, tracking which batches have been reconciled, retried or flagged. Currently it keeps this progress in an in-context scratchpad, and after compaction it sometimes reprocesses batches. What should the context engineer do?

  1. AAsk the model to re-derive which batches are complete by reasoning over the conversation history at each step
  2. BStore the progress in a Delta-backed state table that the agent reads and updates through tools at each step Correct answer
  3. CEnlarge the scratchpad section of the prompt so that the full batch list is preserved verbatim after compaction
  4. DWrite the batch status into the system prompt so that it is the first thing the model reads at every turn

Answer and explanation: Large, exact, frequently updated progress state needs durable storage outside the context window; a Delta-backed table read and written through tools survives compaction and restarts. A larger scratchpad still lives in context and grows with the task. Re-deriving progress from history is unreliable once history is compacted, and rewriting the system prompt every step is costly and fragile.

Question 7 of 10, Memory Architecture with Lakebase and MLflow

A shopping assistant agent stores conversation state in a LangGraph checkpointer keyed by thread ID. Users complain that when they return the next day in a new chat, the agent has forgotten their shoe size and preferred brands that they shared previously. What is the MOST appropriate change?

  1. AAppend the full transcript of every past chat to the system prompt so all earlier preferences are always visible
  2. BIncrease the checkpointer retention period so that the previous day's thread is kept in Lakebase for a longer time
  3. CLog the preferences as MLflow trace tags so that they are available whenever the agent is served again
  4. DPersist preferences in a cross-session Lakebase memory store namespaced by user ID and load them at session start Correct answer

Answer and explanation: Thread-scoped checkpoints hold short-term memory for one conversation; preferences that must survive across sessions belong in a long-term store keyed by user. Keeping the old thread longer does not help because the new chat uses a new thread ID. Replaying every transcript bloats context, and trace tags are observability metadata that the agent does not read at inference time.

Question 8 of 10, Tool Design, MCP, and Agent Context

Why does progressive disclosure of tool information reduce token consumption compared with loading full tool schemas upfront?

  1. AThe model caches all schemas between requests so they never need to be re-sent
  2. BTools are executed on the client, so their descriptions are not needed
  3. CSchemas are converted into embeddings that take up no prompt tokens
  4. DOnly the details relevant to the current step enter the context window Correct answer

Answer and explanation: Progressive disclosure staged loading means the model sees brief summaries first and full schemas only for tools it chooses, so less tool text enters the window per step. Caching is a separate optimisation and does not remove tokens from attention. Schemas are not converted into token-free embeddings, and the model still needs descriptions to call tools correctly wherever they run.

Question 9 of 10, Context Compression and Compaction

A long-running agent drafts a policy report. The team is writing its first compaction prompt and has a set of real traces to test it on. Which tuning approach is recommended?

  1. AMaximise recall first so the summary keeps everything relevant, then iterate on precision Correct answer
  2. BUse a minimal prompt and let the model decide what matters
  3. CStart by maximising precision with a very short summary, then add back items only when a failure is reported in production
  4. DOptimise only for token count, choosing whichever prompt yields the smallest summary across the sample traces

Answer and explanation: The recommended method is recall first, so nothing that later steps need is lost, then precision to remove what is superfluous. Starting with aggressive brevity loses subtle context and learns about it only from production failures. Leaving it entirely to the model and optimising only for size both ignore whether the summary preserves what downstream steps need.

Question 10 of 10, Multi-Agent and Long-Horizon Task Design

A coordinating agent splits a market analysis into three sub-agent tasks by region. Each sub-agent receives only a one-line task message. The final report is inconsistent: one region uses fiscal-year figures and another uses calendar-year figures, even though the coordinator had agreed on fiscal year with the user. What is the root cause?

  1. AThe sub-agents use a different foundation model than the coordinator, which causes them to interpret dates differently
  2. BSub-agents received no shared decisions from the coordinator's trace, so each made its own assumptions Correct answer
  3. CThe coordinator's context window was too large, so it could not hold the outputs of all three sub-agents at once
  4. DRegional data lives in separate Delta tables, preventing consistent results

Answer and explanation: Actions carry implicit decisions; when sub-agents get only a terse task message, they lack the agreed conventions and each fills the gap differently. Model choice and table layout do not explain why a convention agreed with the user was ignored. The coordinator's capacity is not the issue, because the inconsistency originates in the sub-agents' inputs.