Application Development — Agents, Tools and MCP — Databricks Generative AI Engineer practice questions

A supervisor agent occasionally loops between two sub-agents until it times out. The FIRST control to add is

1. an explicit maximum-iteration limit with a defined fallback response
2. switching the supervisor to a model with a larger context window
3. adding provisioned throughput to each of the sub-agent endpoints
4. a nightly batch job that replays the failed conversations for analysis

Answer: A — an explicit maximum-iteration limit with a defined fallback response

A) Correct — bounded iteration with a deterministic fallback is the standard safety control for agent loops. B) A bigger window lets the loop run longer before failing. C) Throughput makes the loop spin faster, not stop. D) Replaying afterwards diagnoses but does not prevent the loop.

When wrapping an internal API as a custom MCP server, which concern should be addressed FIRST?

1. deciding whether tool descriptions should be authored in JSON or in YAML format
2. how the server authenticates to the API and how those secrets are scoped
3. whether the server should be written in Python rather than in Java
4. how many agents will eventually connect to the server over time

Answer: B — how the server authenticates to the API and how those secrets are scoped

B) Correct — the server becomes a privileged bridge, so credential handling, scoping and secret storage are the primary design decisions. A) Serialization detail is trivial by comparison. C) Language choice is an implementation preference. D) Future consumer counts affect scaling, which is addressed after the security boundary is sound.

A developer wants to see, for one failing request, exactly which tools were called, with what arguments, and how long each took. The MOST direct way is to

1. inspect the MLflow trace for that request, which records a span per tool call
2. read the raw driver logs of the cluster that served that particular request
3. query the vector search index's change data feed for the window around the request
4. enable AI Gateway rate limiting and then re-run the failing request by hand

Answer: A — inspect the MLflow trace for that request, which records a span per tool call

A) Correct — tracing is the purpose-built view of an agent's execution: spans, inputs, outputs and timings per step. B) Raw logs are unstructured and lack the span hierarchy. C) The change feed shows index data changes, not agent control flow. D) Rate limiting governs traffic and tells you nothing about a single request's internals.

Function calling in an agent means the LLM:

1. Returns JSON naming a tool and its arguments
2. Directly executes Python functions in its runtime
3. Runs shell commands on the serving host itself
4. Edits system files to persist its own state

Answer: A — Returns JSON naming a tool and its arguments

A) Correct — the model emits structured JSON identifying a tool and arguments; the orchestrator runs it and feeds the result back. B) The agent code executes, never the model. C) The model has no shell. D) Nothing in function calling touches the host filesystem.

A team already runs a vendor-provided MCP server elsewhere in their estate and wants their Databricks agent to use it. Which characterization is correct?

1. external MCP servers must be re-implemented as custom servers before any agent is able to call them
2. an external MCP server is connected as a tool source, and its credentials must still be governed
3. external MCP servers can only be used from batch jobs and never from a served agent endpoint
4. any MCP server hosted outside the workspace is automatically treated as a managed server

Answer: B — an external MCP server is connected as a tool source, and its credentials must still be governed

B) Correct — external servers are a first-class category; you connect them and then handle authentication, secrets and egress under governance. A) Re-implementation defeats the purpose of a shared protocol. C) There is no such batch-only restriction. D) Managed means Databricks-hosted; hosting location is precisely what distinguishes the categories.

A chat app maintains conversation state. Where should the chat history live?

1. Inside the model, which retains prior turns between separate API calls
2. In the vector index only, retrieving past turns by similarity each time
3. On the user's clipboard, pasted back into the app at the start of a turn
4. In application state, replayed into each prompt within the token budget

Answer: D — In application state, replayed into each prompt within the token budget

D) Correct - the application owns the transcript and re-sends the relevant portion, summarizing older turns to stay inside the budget. A) Completion APIs are stateless between calls. B) Similarity search over history loses ordering and recency, so it supplements rather than replaces the transcript. C) Manual pasting is not a state mechanism.

Which risk is MOST specific to persisting agent memory that includes raw user messages?

1. the memory table will eventually exceed Delta Lake's maximum supported column count per table
2. retrieval latency from the memory store will exceed the serving endpoint timeout
3. sensitive personal data is retained beyond its purpose unless masking and retention apply
4. the agent will gradually lose the ability to call its registered tools

Answer: C — sensitive personal data is retained beyond its purpose unless masking and retention apply

C) Correct — durable memory turns transient input into a retained data asset, so masking, access control and retention become mandatory. A) Column limits are irrelevant to an append-only message log. B) A well-indexed keyed read is fast; latency is not the distinguishing risk. D) Memory and tool calling are independent.

A team wants each sub-agent in a supervisor system to be independently versioned and released. Which arrangement BEST supports this?

1. store each sub-agent's prompt text inside the supervisor's own configuration file
2. bundle all of the sub-agents into a single model artifact and version that bundle
3. keep the sub-agents as notebook functions that the supervisor imports at runtime
4. register and serve each sub-agent as its own Unity Catalog model and endpoint

Answer: D — register and serve each sub-agent as its own Unity Catalog model and endpoint

D) Correct — separate registered models and endpoints give each team its own lifecycle, rollback and monitoring boundary. B) One artifact forces lockstep releases, the opposite of the requirement. C) Runtime notebook imports have no versioning or promotion story. A) Prompt strings in a config file are not a deployable, testable agent boundary.

Which statement BEST describes what the LLM actually uses to decide whether to invoke a particular Unity Catalog function as a tool?

1. the function's owner and the set of grants attached to it in Unity Catalog
2. the physical execution plan the SQL warehouse compiles for the function body at call time
3. the function name, parameter names and types, and the comment/description text
4. the row count and update frequency of the tables the function reads from

Answer: C — the function name, parameter names and types, and the comment/description text

C) Correct — the tool schema handed to the model is name, typed parameters and description; a vague comment is the usual cause of a tool never being called. A) Grants control whether the call succeeds, not whether the model chooses it. B) The plan is invisible to the model. D) Data volume never reaches the model's tool-selection step.

Which practice does the MOST to improve a Genie space's answer accuracy?

1. raising the SQL warehouse size so that generated queries finish faster
2. curating a small set of well-described tables and adding example question-and-SQL pairs
3. adding every table in the catalog to the space so that no user question is ever out of scope
4. instructing users to phrase every question as valid SQL before submitting

Answer: B — curating a small set of well-described tables and adding example question-and-SQL pairs

B) Correct — Genie quality is driven by metadata quality and worked examples over a tight table scope. A) Warehouse size affects latency, not correctness. C) A wide scope increases ambiguity and join errors — the opposite of what helps. D) Requiring SQL from users removes the reason to use Genie.

What problem does the Model Context Protocol solve for agent development?

It gives tools and data sources a single standard interface, so any MCP-compatible agent can discover and call them without bespoke per-tool integration code. Tools become reusable across agents and frameworks.

Name the three categories of MCP server available for agent tool integration on Databricks.

Managed MCP servers hosted by Databricks for built-in resources such as Vector Search indexes, UC functions and Genie Spaces • External MCP servers run by third parties • Custom MCP servers you build and host yourself, for example on Databricks Apps.

Why is appending the full conversation history to every agent request a poor persistent-memory strategy?

It grows unbounded, driving up latency, token cost and the risk of exceeding the context window, and it dilutes relevant facts with noise. Persist state in a datastore and retrieve or summarize only the relevant memories per turn.

Mosaic AI Agent Framework?

Build, evaluate, deploy LLM agents using LangGraph or custom Python. Integrated with MLflow tracing + evaluation.