Agentic Solutions — Azure AI Apps and Agents Developer AI-103 practice questions

A team ships an agent whose tool descriptions are terse, such as "gets data". Which consequence is the MOST likely?

1. The tool schema is rejected at registration because descriptions must be verbose
2. The model calls the wrong tool or omits required arguments more often
3. The agent's token cost rises sharply because short descriptions are re-sent often
4. Tool results are returned to the user unformatted rather than summarised

Answer: B — The model calls the wrong tool or omits required arguments more often

B) Correct — descriptions are the model's only guide to what a tool is for and when it applies, so vague text drives mis-selection and malformed argument sets. A) Terse descriptions are syntactically valid; nothing rejects them. C) Short text lowers, not raises, per-turn token cost, so the cost claim is backwards. D) Response formatting is governed by the instructions and the model, not by the length of a tool description.

Users complain that an agent gives wrong answers only in long sessions. The team wants to see exactly which tools ran, with what arguments, in the failing turns. What should they examine FIRST?

1. The agent's traces, which record the turn's tool calls, arguments, and outputs
2. The content filter severity report for the deployment serving the agent
3. The quota and rate-limit metrics for the project's token consumption
4. The evaluation scores for groundedness on the most recent test dataset

Answer: A — The agent's traces, which record the turn's tool calls, arguments, and outputs

A) Correct — tracing captures the per-turn execution span tree, which is the only artefact that shows the actual tool invocations and their arguments and results. B) Content filter reports show safety category hits, not reasoning or tool behaviour. C) Rate-limit metrics explain throttling and latency, not a wrong answer with a successful response. D) Evaluation scores quantify quality in aggregate over a dataset but do not explain an individual failing session.

An agent's inventory tool intermittently returns HTTP 429 from a downstream API. Which handling strategy is MOST appropriate?

1. Retry immediately in a tight loop until the downstream API finally succeeds
2. Treat 429 as fatal and return an apology to the user without any retry
3. Retry with exponential backoff and jitter, then surface a clear failure
4. Swap the failing tool for a different tool that reads from a stale cached file

Answer: C — Retry with exponential backoff and jitter, then surface a clear failure

C) Correct — 429 signals transient throttling, so bounded retries with backoff and jitter recover without stampeding, and a bounded budget stops the turn from hanging forever. A) Tight retry loops amplify the overload and can extend the throttling window. B) Giving up on the first 429 wastes a recoverable condition and hurts success rate. D) Silently substituting stale data trades a visible error for an invisible correctness problem.

An agent answers HR questions from a document set that is updated weekly. Which grounding approach is the MOST appropriate?

1. Fine-tune the base model each week on the updated HR document corpus
2. Give the agent a retrieval tool over an index that is refreshed on the same cadence
3. Paste the full HR corpus into the agent instructions and republish weekly
4. Rely on the base model's pretraining, adding a disclaimer about possible staleness

Answer: B — Give the agent a retrieval tool over an index that is refreshed on the same cadence

B) Correct — retrieval separates volatile knowledge from the model, so refreshing the index updates the agent's answers without touching the model. A) Fine-tuning teaches form and style far more reliably than fresh facts, and a weekly training cycle is slow and costly. C) Instructions have a bounded budget and would be exhausted by a corpus, raising cost on every turn. D) Pretraining cannot contain a private corpus at all, so a disclaimer does not make the answers usable.

A developer is standing up their first hosted agent in Microsoft Foundry and wants the platform to persist conversation state, host the tool definitions, and run the tool-calling loop server side. Which offering is the MOST appropriate starting point?

1. Microsoft Foundry Agent Service, which hosts the agent and its state in the project
2. Azure Machine Learning studio pipelines, orchestrating each model call as a job step
3. A raw chat completion call wrapped in a loop the developer writes and hosts
4. Azure Search in Foundry Tools, using an indexer schedule to drive the turn loop

Answer: A — Microsoft Foundry Agent Service, which hosts the agent and its state in the project

A) Correct — Foundry Agent Service is the managed runtime: the agent definition, its tools, and its conversation state live in the Foundry project, and the service executes the tool-calling loop. B) Azure Machine Learning studio is for training and batch ML workflows, not interactive agent state. C) Hand-rolling the loop is possible but pushes state, retries, and tracing onto the developer, which is exactly what was to be avoided. D) Azure Search in Foundry Tools retrieves grounding data; an indexer is a data-refresh schedule and cannot orchestrate turns.

Which of the following is NOT a legitimate reason to prefer the hosted agent service over a self-managed loop?

1. Conversation state is persisted by the service rather than by custom code
2. Tool execution and tracing are handled by the runtime out of the box
3. Guaranteed factual correctness of every answer the agent produces
4. Consistent integration with project-level identity and role configuration

Answer: C — Guaranteed factual correctness of every answer the agent produces

C) Correct — no runtime can guarantee factual correctness; that remains a grounding, evaluation, and design problem regardless of hosting. A) Managed state removes a substantial amount of custom persistence code. B) Built-in tool orchestration and tracing are genuine benefits of the hosted runtime. D) Sharing the project's identity and role model simplifies secure access to resources.

Which scenario is the LEAST appropriate candidate for a multi-agent design?

1. A single-step FAQ lookup answered from one index with no branching logic
2. A travel workflow needing separate flight, hotel, and expense-policy specialists
3. A review pipeline where one agent drafts and a second critiques the draft
4. A workflow whose steps require distinct tool permissions and separate identities

Answer: A — A single-step FAQ lookup answered from one index with no branching logic

A) Correct — a single retrieval step over one index gains nothing from orchestration and pays extra latency, cost, and failure surface for the privilege. B) Distinct domains with different tools and knowledge are the canonical case for specialists. C) Drafter and critic roles benefit from separation because each has a different objective. D) Splitting by permission boundary lets each agent hold only the rights its own step requires.

Which statement about how a model decides to call a tool is accurate?

1. The model selects a tool from the supplied schemas based on the request and context
2. Tools fire in the order they were registered, once per turn, regardless of intent
3. The runtime pattern-matches keywords in the user text and picks the tool entirely on its own
4. A tool is invoked only when the user names it explicitly in their own message

Answer: A — The model selects a tool from the supplied schemas based on the request and context

A) Correct — the model reasons over the tool names, descriptions, and parameter schemas presented to it and emits a structured call when one fits the request. B) Registration order is not an execution schedule, and a turn may call zero, one, or several tools. C) There is no keyword matcher in front of the model; selection is a model decision. D) Requiring an explicit user mention would defeat the point of tool calling, though a tool-choice setting can force or forbid specific tools.

What does a function (tool) call return to the calling application?

1. A synthesized audio rendering of the function's arguments for playback
2. Nothing to return, because the service runs the named function for you
3. An image describing the tool schema that the developer registered earlier
4. A JSON argument payload the app runs, then feeds the result back to model

Answer: D — A JSON argument payload the app runs, then feeds the result back to model

D) Right — the model emits the function name and a JSON argument object; your code executes it and returns the result as a new item for the model to use. A) Tool calls are structured text, not audio. B) The service never executes your code. C) Schemas are supplied by you, not rendered as images.

A developer wants an agent's earlier turns to influence later answers within the same session, without resending the whole history manually on each call. Which mechanism provides this?

1. Increasing the max output tokens on the deployment used by the agent
2. Enabling semantic ranking on the search index backing the retrieval tool
3. Setting a higher top-p so the model samples from more of the distribution
4. Persisting the exchange in a conversation the service replays into each response

Answer: D — Persisting the exchange in a conversation the service replays into each response

D) Correct — the conversation object is the durable container of prior items, and the service supplies them to the model when it creates the next response. A) Output token limits govern reply length, not what history the model sees. B) Semantic ranking reorders retrieved documents; it has nothing to do with session memory. C) Sampling parameters change wording variety, not recall of earlier turns.

What are the practical limits of conversation memory, and how are they managed?

Everything the model can use must fit the context window, and every retained turn costs tokens on every subsequent call. Manage it by summarizing older turns, pinning the durable facts, and moving long-lived knowledge into retrieval rather than keeping raw history.

What does a tool schema declare, and why is the description field load-bearing?

It declares the tool name, a natural-language description, and a typed parameter schema with required fields and enumerations. The model chooses tools by reading descriptions, so an ambiguous or overlapping description is the most common cause of the wrong tool being called.

What is the Microsoft Agent Framework and how does it relate to Foundry Agent Service?

It is the open-source SDK for building and orchestrating agents in code — defining agents, registering tools, composing workflows and multi-agent patterns. It can drive locally hosted models or target agents hosted in Foundry Agent Service, which supplies the managed runtime, state and governance.

What is Microsoft Foundry Agent Service?

It is the managed Azure service that hosts agents: it holds the agent definition — model, instructions and tool set — manages conversation state, executes the tool-calling loop, and applies the project's identity, networking and content-safety configuration.