A context engineer is developing a multi-agent research pipeline on Databricks. A coordinator agent divides an incoming scientific literature query into two parallel subtasks: a citation-retrieval agent and a methodology-analysis agent. Both agents independently receive only the original user query, produce their outputs, and return them to the coordinator, which then invokes a synthesis agent. The synthesis agent produces a final report that contradicts the citation-retrieval findings because it relies solely on the methodology-analysis output and has no visibility into retrieved citations.
Which two context propagation changes would prevent the conflicting outputs in the synthesis agent?
Choose two
A context engineer is developing a RAG pipeline for a pharmaceutical knowledge base. Document chunks from clinical trial summaries are stored in a Unity Catalog-governed Delta table with precomputed embedding vectors. Pure vector similarity search retrieves topically related chunks effectively but misses chunks that depend on precise drug compound names and dosage identifiers in the search string, because those tokens are underrepresented in the embedding space. The engineer wants to recover those exact-term chunks while preserving the semantic retrieval that already works, without standing up a separate keyword index or adding a downstream merge step.
Which configuration should the engineer use to close this retrieval gap?
A context engineer is building a regulatory compliance agent on Databricks that connects to an external compliance management platform through an MCP Service registered in Unity Catalog.
The MCP server exposes the following tools:
search_compliance_records: Retrieves existing compliance records without modifying them.
get_submission_status: Checks the status of previously submitted reports.
submit_regulatory_report: Sends a finalized regulatory report to an external authority.
delete_compliance_record: Permanently deletes an existing compliance record.
The agent serves compliance officers who are authorized to retrieve records and submit reports. However, organizational policy requires explicit human approval before every regulatory report submission.
The agent must continue executing read-only tools without interruption. It must also prevent unauthorized destructive operations and maintain an auditable record of tool invocations.
The engineering team has confirmed that the MCP tools function correctly. The remaining requirement is to enforce the approval and authorization rules at the platform governance layer rather than relying exclusively on instructions inside the agent ' s system prompt.
Which approach should the context engineer implement?
A context engineer is examining an MLflow trace for a Databricks compliance agent mid-workflow on a quarterly regulatory review for a regional bank. The agent’s design pattern is to retrieve regulatory clauses, extract structured obligations with regional metadata and source citations, validate each obligation through the rules engine, and synthesize the audit memo from the structured intermediate plus the validation outcomes. The context window is nearing the compaction threshold, and compaction will fire before the workflow is complete.
The trace includes three executed spans and one queued span:
Span 1 (RETRIEVER): Retrieved 22 regulatory clause excerpts from the compliance vector index. Output: 22 documents (8,400 tokens) of raw clause text with metadata for source URI and section ID.
Span 2 (LLM/CHAT_MODEL): Produced a structured intermediate (410 tokens) containing three extracted compliance obligations, each annotated with a regional applicability tag describing how the bank’s state-level regulations modify the federal requirement and a source citation pointing back to the originating clause.
Span 3 (TOOL): Ran the three obligations through the rules engine and returned a structured pass/fail report (280 tokens) with one outcome entry per obligation. Each entry contains the obligation ID, the pass/fail value, and a brief explanation of the rule that determined the outcome.
Span 4 (LLM/CHAT_MODEL, queued): Will draft the compliance memo for the bank’s audit team.
Which content from the executed spans can the compactor safely summarize without affecting span 4’s execution?
A context engineer maintains a Databricks-based assistant for a fraud investigation team. Each investigation is a long-running thread that the team works on across many sessions over weeks or months. Investigations are typically handled by two or three analysts who rotate in and out as new evidence arrives, and any one analyst is usually working several open investigations at the same time. The assistant runs on Agent Bricks Custom Agents with a single retrieval tool over a Databricks AI Search index of the team’s case file corpus (incident reports, suspect dossiers, transaction logs, exhibits), synced nightly from the source Delta tables that the case management system writes to. The retrieval tool is called with num_results=50. The system has been in production for nine months.
The team begins reporting that the assistant treats every new session as a cold start. The most recent example: on Friday, analyst A and the assistant work together on an active investigation. By the end of the session, they have eliminated two suspects on alibi evidence, confirmed a third as the likely actor, and identified two specific accounts as the transfer pathway for the funds in question. On Monday, analyst B opens the assistant against the same investigation to draft the next subpoena request. The assistant responds by asking who the suspect is, what the eliminations were, and which accounts are in scope. The engineer reproduces the pattern across other investigations: within a single session, the assistant reasons well; the next session, regardless of which analyst opens it, starts cold.
What change should the engineer make to address this pattern?