Task 2.2: Build agents by using Foundry

• AAdd the earlier turns to the agent's instructions after each reply

  Incorrect

  Instructions define the agent for all users; rewriting them per chat would mix users' histories.

• BSend all earlier messages in every request and start a new conversation

  Incorrect

  That works but is the resending the app wants to avoid, and it grows token use every turn.

• CCreate a conversation and pass its ID with every request to the agent

  Correct

  A conversation keeps the history of turns, so each request only adds the new user message.

• DSend each follow-up with a new conversation ID

  Incorrect

  A new conversation starts without the earlier turns, so follow-ups lose their context.

Conversation state is the turn-by-turn history of a chat: create a conversation once and reference it on each turn instead of resending history.

• ARemove web search and rely on the model's own knowledge for HR answers

  Incorrect

  That drops the fallback the team wants and swaps web results for the model's training data.

• BRaise the model's temperature so that it explores the other tool more

  Incorrect

  Randomness doesn't encode a preference for internal content.

• CGive both tools identical, generic descriptions

  Incorrect

  Generic descriptions make it harder for the model to tell the tools apart, not easier.

• DAdd an instruction to use file search before web search for internal topics

  Correct

  When tools overlap, the guidance is to tell the model what each tool is for and add a decision rule such as this one.

Tool choice is driven by instructions and tool descriptions. With overlapping tools, state each tool's purpose and an explicit order of use.

• ARun the function and send the result as function_call_output for that call_id

  Correct

  Your app executes the function and submits the output for that call; the agent then uses it to finish its answer.

• BUpload the tracking data as a code interpreter file and ask the question again

  Incorrect

  The tracking system lives in your network; a file upload isn't a live lookup of the parcel's status.

• CWait for the agent service to call the tracking system on its own behalf

  Incorrect

  With a function tool the service never runs your code; it only asks your app to.

• DStart a new conversation that includes the shipment status as plain text

  Incorrect

  A new conversation loses the pending call and the context of the user's question.

Function calling is a round trip: the model asks for a call, your app runs it and returns the output tied to the call_id, and the agent continues.

• AAn OpenAPI tool with API key authentication stored in a project connection

  Incorrect

  That keeps a secret in a connection, and the API doesn't accept keys.

• BThe web search tool restricted to the API's host name with a site filter

  Incorrect

  Web search queries public web results; it can't call a private API's operations.

• CAn OpenAPI tool with anonymous authentication and a private endpoint

  Incorrect

  The API rejects calls that carry no Entra token.

• DAn OpenAPI tool that uses managed identity auth for the project's identity

  Correct

  OpenAPI tools support anonymous, API key and managed identity auth; managed identity grants the project's identity a role on the API and stores no secret.

Existing HTTP APIs with an OpenAPI 3.0 or 3.1 description plug in as OpenAPI tools. Choose the auth method the API accepts; for Entra-protected Azure APIs with no secrets allowed, that is managed identity.

• AA group chat orchestration in Microsoft Agent Framework

  Incorrect

  In group chat, agents collaborate in one shared conversation; here one specialist at a time takes over the customer.

• BA new workflow in the Foundry portal's visual workflow builder

  Incorrect

  Foundry is retiring workflows on December 1, 2026, and directs new development to Microsoft Agent Framework.

• CA handoff orchestration in Microsoft Agent Framework

  Correct

  Handoff is the built-in pattern in which agents transfer control to each other based on context.

• DThe Connected Agents tool on the triage agent

  Incorrect

  Connected Agents belongs to the classic Agents API and isn't available in the new Foundry Agent Service.

Agent Framework has several built-in orchestration patterns. A conversation that moves between specialists based on context is a handoff.

• AAsk the model, in its instructions, to confirm with the lead before each refund

  Incorrect

  An instruction isn't an enforced control; the model can still call the function.

• BSet tool_choice to none whenever the user's message mentions a refund

  Incorrect

  That blocks every tool call in those turns, including the free ones, and never asks the lead.

• CRemove the refund function and email the lead the details of each refund

  Incorrect

  The agent loses the ability to complete refunds after approval.

• DWrap the refund function in ApprovalRequiredAIFunction and resend the decision

  Correct

  The run then completes with a request for approval instead of executing; the caller collects the decision and passes it back.

Human approval belongs in the tool definition, not the prompt. In Agent Framework, functions marked approval-required pause the run and return an approval request that your app answers in a new run.

• ACoherence

  Incorrect

  Coherence rates the logical flow of the reply, not what happened in the tool step.

• BTool Call Success

  Correct

  Tool Call Success evaluates whether tool calls executed without technical failures.

• CIntent Resolution

  Incorrect

  Intent Resolution checks whether the agent understood what the user wanted; here it did, and the tool step is what breaks.

• DTool Input Accuracy

  Correct

  Tool Input Accuracy checks that call parameters are correct in grounding, type, format and completeness.

• ESimilarity

  Incorrect

  Similarity compares response text with a reference answer and can't see tool calls.

Agent evaluators look inside the run. Tool Call Success isolates execution failures, Tool Input Accuracy isolates wrong arguments, and together they tell you which side of the tool call to fix.

• AConcurrent orchestration of the three specialist agents

  Incorrect

  No manager builds a plan here or changes it as findings come in.

• BSequential orchestration of the three specialist agents

  Incorrect

  No manager tracks progress here or replans when an approach stalls.

• CGroup Chat orchestration with a round-robin manager

  Incorrect

  A round-robin manager picks speakers in turn; it doesn't plan, track progress or replan the way the Magentic manager does.

• DMagentic orchestration with a manager agent that plans

  Correct

  Magentic suits tasks whose solution path isn't known in advance; its manager plans, delegates, tracks progress and adapts the plan.

Magentic has the same architecture as Group Chat, with a manager that uses planning to coordinate the agents. If the scenario needs simpler coordination without complex planning, Group Chat is the lighter choice.

• AGroundedness

  Incorrect

  It judges the final answer, not which tools were chosen.

• BSimilarity

  Incorrect

  Similarity compares a response with a reference text, not the tools the agent chose.

• CTool Selection

  Correct

  Tool Selection measures whether the agent selected the most appropriate and efficient tools for a task.

• DCoherence

  Incorrect

  Coherence rates the logical flow of replies.

Agent evaluators split tool use into several separate checks. Pick the one that matches the failure you suspect.

• AFunction calling middleware registered on the agent

  Correct

  Function calling middleware intercepts all function calls the agent executes so input and output can be inspected and modified.

• BAgent run middleware that logs each run's input and output

  Incorrect

  Agent run middleware wraps whole runs; it can't change each function's result before the model reads it.

• CA checkpoint manager on the agent's workflow

  Incorrect

  It belongs to workflows and doesn't intercept the function calls an agent makes.

• DA rewrite of every function to log and mask its own data

  Incorrect

  That is the per-function editing the team wants to avoid.

Agent Framework middleware comes in agent run, function calling and chat client flavors, chained so each calls the next. Pick the layer that sees the data you need to change.

• ARaise the tool's timeout so the agent waits quietly for the result

  Incorrect

  A longer silent wait is the dead air callers complain about; the guidance is a short spoken progress response when a call takes time.

• BReturn only needed facts and speak a short progress line

  Correct

  For voice agents, keep tool output to what the turn needs and give a short spoken progress response when a call takes time.

• CReturn the full JSON payload so that the agent can read every field

  Incorrect

  Long payloads lead to long spoken answers; return only what the agent needs for the current request.

• DGive the tool a longer description that asks for detailed answers

  Incorrect

  The guidance keeps tool descriptions concise, and detailed answers make the spoken replies even longer.

Tool latency and response length shape a caller's experience. Design tools for voice with concise descriptions, small results and graceful handling of interruptions and empty results.

• ACreate an agent session and pass it to each run

  Correct

  An agent session holds conversation state, so runs in the same session see earlier turns.

• BRaise the model's max output tokens for each of the runs

  Incorrect

  Output length doesn't preserve history.

• CAdd a checkpoint manager to the agent's run calls

  Incorrect

  It doesn't hold chat history between an agent's runs.

• DCreate a new agent object for every message in the chat

  Incorrect

  New agents start without the earlier turns.

In Agent Framework, sessions carry state between runs and context providers can enrich them; both are part of the framework's building blocks.