I85PREP
⌘ K
AI-500 study materials
AI-500 study resources

AI-500 practice questions

One question at a time. Build your AI-500 understanding.

First 50 of 200 questions. Each one lists the choices, the correct answer, and an explanation. The checker at the top of this page saves progress in this browser when it is available.

Continue in the practice tool above to load more questions, 50 at a time.

Question 1Architect multi-agent solutions

In a multi-agent architecture, which component is primarily responsible for decomposing a user goal into subtasks and assigning them to specialized agents?

  1. The guardrail service
  2. The retrieval agent
  3. The orchestrator agent Correct answer.
  4. The message broker

Explanation

An orchestrator (or supervisor) agent owns planning and delegation: it breaks the goal into subtasks and routes them to the appropriate specialist agents. Retrieval agents fetch knowledge, guardrail services enforce policy, and brokers move messages.

Question 2Architect multi-agent solutions

Two agents must exchange structured results without either knowing the other's network location. Which design element most directly provides this decoupling?

  1. A shared message bus with topic-based routing Correct answer.
  2. A hardcoded direct HTTP call from one agent to the other
  3. A shared in-memory variable inside one process
  4. A single monolithic prompt containing both agents' instructions

Explanation

A message bus with topic-based routing lets producers publish and consumers subscribe without knowing each other's addresses, which is the definition of loose coupling. Direct calls and shared memory both create tight dependencies.

Question 3Architect multi-agent solutions

An agent repeatedly calls the same tool with identical arguments and receives identical results, consuming the step budget. Which architectural control best addresses this?

  1. A per-agent loop-detection rule that halts or escalates on repeated identical calls Correct answer.
  2. Removing the tool from the agent's toolset entirely
  3. Adding more specialized agents to the pipeline
  4. Increasing the model temperature for that agent

Explanation

Loop detection compares successive tool invocations and interrupts cycles, preserving budget and surfacing the problem. Raising temperature does not stop repetition, adding agents increases cost, and removing the tool destroys needed capability.

Question 4Architect multi-agent solutions

Which statement best describes the supervisor pattern in a multi-agent system?

  1. All agents share one prompt and execute in a single step
  2. Worker agents negotiate directly with each other without a coordinator
  3. A human approves every individual tool call
  4. A central agent decides which worker agents run and consolidates their outputs Correct answer.

Explanation

In the supervisor pattern a coordinating agent selects workers, sequences their execution, and merges results. Peer negotiation describes decentralized designs, and human approval of every call describes a manual gate, not a supervisor.

Question 5Architect multi-agent solutions

A team wants each specialist agent to be independently deployable and versioned while still participating in one workflow. Which architecture most directly supports this?

  1. A single agent with a growing list of tools and no separation
  2. Each agent exposed as an independently versioned service invoked through the orchestrator Correct answer.
  3. A shared database that all agents read and write directly
  4. One large prompt template containing all specialist instructions

Explanation

Service-per-agent boundaries allow independent deployment, scaling, and versioning while the orchestrator composes them into a workflow. Monolithic prompts and shared mutable databases couple the agents tightly.

Question 6Architect multi-agent solutions

What is the primary purpose of assigning a narrowly scoped role and toolset to each agent in a multi-agent system?

  1. To guarantee the system never produces an incorrect answer
  2. To reduce ambiguity in tool selection and improve reliability of each agent's decisions Correct answer.
  3. To eliminate the need for any orchestration layer
  4. To make each agent's context window unlimited

Explanation

Narrow scope shrinks the decision space, making tool selection and reasoning more predictable and easier to evaluate. It does not guarantee correctness, remove orchestration, or expand context limits.

Question 7Architect multi-agent solutions

An orchestrator must combine findings from three research agents whose outputs partially conflict. Which approach best supports a defensible consolidated answer?

  1. Have the orchestrator reconcile conflicts explicitly and cite which agent supports each claim Correct answer.
  2. Use whichever agent responded fastest
  3. Average the three outputs into a single blended paragraph
  4. Discard all three outputs and answer from the orchestrator's own knowledge

Explanation

Explicit reconciliation with provenance lets the orchestrator surface disagreements and justify the final synthesis. Blending, speed-based selection, and discarding evidence all hide or lose the basis for the answer.

Question 8Architect multi-agent solutions

Which is the most appropriate use of a handoff between two agents?

  1. Replacing the orchestrator's planning responsibility
  2. Duplicating the same task across agents for redundancy on every request
  3. Avoiding the need to define any shared context schema
  4. Transferring control when the current agent lacks the capability or authority to continue Correct answer.

Explanation

Handoffs exist to move control to an agent with the right capability or authority. Using them for redundant duplication, to skip schema design, or to replace planning misuses the mechanism.

Question 9Architect multi-agent solutions

What does a shared context or state object passed between agents primarily enable?

  1. Consistent task state and intermediate results across agent boundaries Correct answer.
  2. Faster model inference per agent
  3. Automatic reduction of token cost to zero
  4. Elimination of the need for access control

Explanation

A shared state object carries task progress and intermediate artifacts so downstream agents can act on prior work. It does not change inference speed, eliminate cost, or remove access-control needs.

Question 10Architect multi-agent solutions

A workflow requires a strict sequence where each step depends on the prior step's validated output. Which orchestration style fits best?

  1. A free-form debate among agents until consensus emerges
  2. Random selection of the next agent at each turn
  3. Parallel fan-out of all steps at once
  4. A deterministic pipeline with explicit validation gates between steps Correct answer.

Explanation

Dependent steps with validation requirements call for deterministic sequencing with gates that block progression on failure. Debate, parallel fan-out, and random selection all violate the dependency ordering.

Question 11Architect multi-agent solutions

Which scenario most clearly justifies a multi-agent design over a single agent?

  1. The task must run with the smallest possible number of model invocations
  2. The task is a single short factual lookup
  3. The task requires distinct specialties with different tools and separate access permissions Correct answer.
  4. The task requires only one tool call

Explanation

Separate specialties, toolsets, and permission boundaries are the classic justification for decomposing into multiple agents. Simple lookups and minimal-invocation goals favor a single agent.

Question 12Architect multi-agent solutions

Which mechanism best enforces that a worker agent cannot invoke tools outside its approved scope?

  1. Increasing the agent's maximum token limit
  2. Logging tool calls after they complete
  3. A policy layer that validates each tool call against the agent's declared permissions Correct answer.
  4. Relying on the model's system prompt alone

Explanation

Enforcement requires an external policy layer that checks each call against declared permissions before execution. Prompt instructions are advisory, post-hoc logging does not prevent action, and token limits are unrelated to authorization.

Question 13Architect multi-agent solutions

What is the main risk of allowing agents to recursively spawn new agents without a depth limit?

  1. Unbounded fan-out that multiplies cost and can exhaust quotas Correct answer.
  2. Automatic improvement in answer accuracy
  3. Reduced need for observability
  4. Improved determinism of the overall workflow

Explanation

Recursive spawning without a depth or count cap can explode the number of active agents, driving cost and quota exhaustion. It does not improve determinism or accuracy and increases observability needs.

Question 14Architect multi-agent solutions

Which practice best supports debugging a multi-agent workflow that produced an incorrect final answer?

  1. Reviewing only the final agent's output
  2. Disabling logging to reduce noise
  3. Correlating a shared trace ID across every agent and tool invocation Correct answer.
  4. Restarting the workflow until a correct answer appears

Explanation

A shared trace ID links each agent and tool step into one causal chain, making the failure point identifiable. Reviewing only the last step, retrying blindly, and disabling logs all obscure the root cause.

Question 15Architect multi-agent solutions

An orchestrator sends the same subtask to two agents and selects the better result. Which trade-off does this design introduce?

  1. Lower cost with no change in latency
  2. Elimination of the need for result evaluation
  3. Higher cost and latency in exchange for improved output quality or resilience Correct answer.
  4. Guaranteed correctness with no additional resource use

Explanation

Redundant execution consumes additional model and tool calls, raising cost and latency, in return for quality or resilience gains. It does not reduce cost, guarantee correctness, or remove the need to evaluate results.

Question 16Architect multi-agent solutions

Which describes a correct use of an agent registry in a multi-agent system?

  1. Agents discover available collaborators and their capabilities at runtime Correct answer.
  2. Agents store their private chain-of-thought for permanent retention
  3. Agents bypass the orchestrator for all decisions
  4. Agents share raw credentials with each other

Explanation

A registry publishes agent identities, capabilities, and endpoints so collaborators can be discovered dynamically. It is not a store for private reasoning, a bypass for orchestration, or a credential-sharing mechanism.

Question 17Architect multi-agent solutions

A regulated workflow requires that no agent act on data outside its assigned jurisdiction. Which architectural control enforces this most reliably?

  1. Instructions in the orchestrator prompt asking agents to stay in scope
  2. Per-agent data scoping enforced at the data access layer Correct answer.
  3. A larger context window for each agent
  4. Periodic manual audits after processing

Explanation

Enforcement at the data access layer constrains what each agent can retrieve regardless of its prompt. Prompt instructions are advisory, audits are retrospective, and context size is unrelated to jurisdiction.

Question 18Architect multi-agent solutions

Which pattern best handles a task where independent subproblems can be solved simultaneously and merged afterward?

  1. A strict sequential chain of agents
  2. Parallel fan-out to worker agents followed by a merge step Correct answer.
  3. A debate loop with no termination condition
  4. A single agent retrying until it succeeds

Explanation

Independent subproblems map naturally to parallel fan-out with a merge stage that combines results. Sequential chains serialize unnecessarily, and retry or unterminated debate loops do not exploit independence.

Question 19Architect multi-agent solutions

What is the key benefit of defining an explicit message schema between agents?

  1. Guaranteed lower model latency
  2. Removal of the need for error handling
  3. Predictable parsing and validation of inter-agent payloads Correct answer.
  4. Automatic translation into every human language

Explanation

A schema makes payloads machine-validatable and predictable, reducing integration failures. It does not affect model latency, remove error handling, or provide translation.

Question 20Architect multi-agent solutions

An agent's tool call fails with a transient timeout. Which orchestration behavior is most appropriate?

  1. Immediately fail the entire workflow with no retry
  2. Retry with bounded exponential backoff and escalate after the limit Correct answer.
  3. Silently substitute a fabricated result
  4. Retry indefinitely until the call succeeds

Explanation

Bounded retries with backoff handle transient faults while capping cost and latency, and escalation prevents silent stalls. Infinite retries risk hangs, immediate failure wastes recoverable errors, and fabrication corrupts outputs.

Question 21Architect multi-agent solutions

Which design most effectively prevents one agent's hallucinated output from being treated as verified fact by downstream agents?

  1. Requiring downstream agents to validate claims against an authoritative source before acting Correct answer.
  2. Removing validation steps to reduce latency
  3. Passing the output along with a label marking it as trusted
  4. Increasing the upstream agent's temperature

Explanation

Independent validation against an authoritative source is what converts a claim into trusted input. Labeling it trusted, raising temperature, or removing validation all increase the risk of propagating errors.

Question 22Architect multi-agent solutions

Which is the primary reason to cap the number of turns an agent may take in a workflow?

  1. To eliminate the need for orchestration
  2. To guarantee the agent reaches a correct conclusion
  3. To increase the model's reasoning depth
  4. To bound cost and latency and force a defined fallback path Correct answer.

Explanation

Turn caps bound resource consumption and guarantee the workflow exits through a defined fallback. They do not increase reasoning depth, guarantee correctness, or replace orchestration.

Question 23Architect multi-agent solutions

A supervisor agent must choose among specialist agents whose capabilities overlap. Which approach yields the most reliable routing?

  1. Routing to whichever agent was created first
  2. Routing on declared capability metadata and required permissions Correct answer.
  3. Routing randomly to distribute load
  4. Routing all requests to every agent simultaneously

Explanation

Routing based on declared capabilities and required permissions matches the task to an agent that can legally and technically handle it. Creation order and random routing ignore fitness, and broadcasting to all agents inflates cost.

Question 24Architect multi-agent solutions

Which statement about human-in-the-loop checkpoints in a multi-agent workflow is accurate?

  1. They are required before every single tool call
  2. They guarantee the workflow cannot fail
  3. They should be placed at high-impact decision points rather than at every step Correct answer.
  4. They eliminate the need for automated guardrails

Explanation

Checkpoints add value where decisions carry significant risk or irreversibility, not at every step, which would defeat automation. They complement rather than replace automated guardrails and cannot guarantee success.

Question 25Architect multi-agent solutions

What is the main purpose of assigning distinct identities and credentials to each agent?

  1. To increase the model's context window
  2. To enable least-privilege access and attributable auditing Correct answer.
  3. To reduce the number of agents needed
  4. To allow agents to share one universal secret

Explanation

Distinct identities let each agent receive only the permissions it needs and make actions attributable in audit logs. They do not affect context size, reduce agent count, or justify shared secrets.

Question 26Architect multi-agent solutions

Which statement best describes the role of a supervisor agent in a hierarchical multi-agent architecture?

  1. It replaces the need for any message contract between agents by sharing raw memory.
  2. It decomposes the overall task and routes subtasks to specialist agents based on their declared capabilities. Correct answer.
  3. It executes every subtask itself and uses other agents only as passive data stores.
  4. It guarantees deterministic output ordering regardless of agent concurrency.

Explanation

A supervisor coordinates work: it interprets the goal, splits it into subtasks, and dispatches those subtasks to agents whose capabilities match. It does not perform all work itself, and it does not remove the need for explicit message contracts.

Question 27Architect multi-agent solutions

An architect wants agents to be discoverable by capability rather than by hard-coded names. Which design element most directly supports this?

  1. A longer maximum turn limit for each agent.
  2. A capability descriptor published with each agent's registration record. Correct answer.
  3. A retry policy applied to all tool calls.
  4. A single shared credential used by every agent.

Explanation

Capability-based discovery requires each agent to advertise what it can do, typically through a descriptor stored in a registry. Callers then query by capability instead of referencing fixed agent identifiers.

Question 28Architect multi-agent solutions

Which message-passing style most reduces coupling between a producer agent and a consumer agent?

  1. Direct in-process function calls with shared mutable objects.
  2. Passing pointers to the producer's internal state.
  3. Asynchronous delivery through a durable queue that the consumer polls. Correct answer.
  4. A synchronous call that blocks until the consumer replies.

Explanation

A durable queue decouples the two agents in time and identity: the producer does not need the consumer to be running or reachable. Shared mutable objects and internal-state pointers create tight coupling and hidden dependencies.

Question 29Architect multi-agent solutions

Which property of an agent's role definition most helps a reviewer predict what that agent will do at runtime?

  1. A written statement of the agent's goal, allowed tools, and stopping conditions. Correct answer.
  2. The physical server on which the agent is deployed.
  3. The number of other agents currently registered.
  4. The programming language used to implement the agent.

Explanation

A role definition that names the goal, permitted tools, and stop conditions gives a reviewer a bounded contract to check behavior against. Implementation language, host, and registry size do not describe behavior.

Question 30Architect multi-agent solutions

A planner agent produces a task graph where two nodes have no dependency between them, but the executor runs them strictly one after the other. What is the main consequence?

  1. The two nodes will produce conflicting outputs.
  2. The executor loses the ability to retry failed nodes.
  3. Available parallelism is left unused, increasing end-to-end latency. Correct answer.
  4. The task graph becomes invalid and must be regenerated.

Explanation

Independent nodes in a task graph are candidates for concurrent execution. Running them serially is still correct but forgoes the parallelism the graph permits, which lengthens total runtime.

Question 31Architect multi-agent solutions

A specialist agent returns a result with a confidence score, and the supervisor must decide whether to accept it or escalate to a second agent. Which policy is most defensible?

  1. Never escalate, because the first agent has the most context.
  2. Escalate only when the result is longer than a fixed character count.
  3. Escalate when confidence falls below a threshold calibrated against known-good outcomes. Correct answer.
  4. Always escalate, because a second opinion is never harmful.

Explanation

A threshold tied to observed outcomes turns confidence into an actionable decision rule with measurable precision and recall. Unconditional escalation wastes resources, and unconditional acceptance ignores known failure modes.

Question 32Architect multi-agent solutions

Two agents must jointly edit a shared document object. Which coordination mechanism best prevents one agent from overwriting the other's changes?

  1. Logging every write to a central audit trail.
  2. Assigning both agents the same read-only toolset.
  3. Increasing each agent's maximum turn count.
  4. Optimistic concurrency control using a version tag checked on write. Correct answer.

Explanation

Version-tagged writes let an agent detect that the object changed since it read it, so conflicting updates are rejected and can be retried. Audit logging records what happened but does not prevent the overwrite.

Question 33Architect multi-agent solutions

An architect is choosing between a fixed pipeline of agents and a dynamic planner for a workflow whose steps are known in advance and rarely change. Which choice is better justified?

  1. The dynamic planner, because planning always produces higher-quality results.
  2. The dynamic planner, because it removes the need for agent role definitions.
  3. The fixed pipeline, because it eliminates the need for any error handling.
  4. The fixed pipeline, because predictable steps favor simpler, cheaper, and more testable orchestration. Correct answer.

Explanation

When the sequence of steps is stable, an explicit pipeline is easier to test, cheaper to run, and simpler to reason about than a planner that re-derives the same steps each time. Planning adds value mainly when steps must be discovered at runtime.

Question 34Architect multi-agent solutions

A supervisor fans a query out to several specialist agents and then merges their answers. Which design detail most improves the quality of the merged result?

  1. Having each agent summarize the others' outputs before replying.
  2. Requiring each agent to return structured findings with explicit provenance. Correct answer.
  3. Allowing each agent to return free-form prose of any length.
  4. Sending the query to the same agent multiple times.

Explanation

Structured findings with provenance let the merger align claims, detect conflicts, and weight sources. Free-form prose makes automated comparison unreliable, and having agents summarize each other introduces cross-contamination.

Question 35Architect multi-agent solutions

A multi-agent system must remain functional when one worker agent becomes unresponsive. Which architectural property addresses this most directly?

  1. A longer timeout on the initial user request.
  2. Health checks combined with supervisor reassignment of the subtask. Correct answer.
  3. Stricter output formatting rules for each agent.
  4. A larger shared context object for all agents.

Explanation

Detecting an unresponsive worker through health checks and reassigning its subtask lets the system route around the failure. Formatting rules and larger context do not restore lost capacity.

Question 36Architect multi-agent solutions

An architect observes that agents in a mesh topology repeatedly re-derive the same intermediate facts, inflating cost. Which change most directly addresses the root cause?

  1. Increase each agent's token budget.
  2. Switch all inter-agent messages to a binary encoding.
  3. Add more agents to the mesh to spread the workload.
  4. Introduce a shared, versioned artifact store that agents read before recomputing. Correct answer.

Explanation

Duplicated derivation usually stems from the absence of a common place to publish and retrieve intermediate results. A versioned artifact store lets agents reuse prior work instead of recomputing it.

Question 37Architect multi-agent solutions

A team wants to evaluate whether a proposed multi-agent decomposition actually improves outcomes over a single capable agent. Which evaluation design gives the most trustworthy answer?

  1. Measure only the multi-agent system's total token consumption.
  2. Compare the multi-agent system's best run against the single agent's average run.
  3. Ask each agent in the multi-agent system to rate its own contribution.
  4. Run both designs on the same held-out task set and compare quality, cost, and latency under identical budgets. Correct answer.

Explanation

A fair comparison holds the task set and resource budget constant and measures several outcomes, so differences can be attributed to the architecture rather than to uneven conditions. Self-assessment and single metrics do not establish comparative benefit.

Question 38Develop multi-agent solutions in Azure

In a multi-agent solution built on the Azure AI Agents service, which component is responsible for coordinating the order in which individual agents are invoked?

  1. The orchestration layer of the client application Correct answer.
  2. The Azure Key Vault instance
  3. The individual agent's model deployment
  4. The storage account holding thread history

Explanation

The orchestration layer decides which agent runs, in what sequence, and how outputs flow between agents. Agents themselves only handle a single turn of reasoning against their configured model and tools.

Question 39Develop multi-agent solutions in Azure

Which Azure AI Agents concept represents a durable, server-side conversation history that multiple runs can append to?

  1. A run
  2. A thread Correct answer.
  3. A vector index
  4. A tool

Explanation

A thread is the persistent conversation container to which messages are added and against which runs execute. Runs are individual executions, and tools are capabilities an agent can call.

Question 40Develop multi-agent solutions in Azure

You are designing a multi-agent system where a triage agent must hand a request to a billing specialist agent only when the request concerns invoices. Which pattern best describes this design?

  1. Reflection pattern
  2. Fan-out/fan-in pattern
  3. Router pattern Correct answer.
  4. Broadcast pattern

Explanation

A router classifies an incoming request and dispatches it to the single most appropriate specialist. Fan-out and broadcast send work to multiple agents, which is not what conditional handoff requires.

Question 41Develop multi-agent solutions in Azure

A specialist agent needs to look up a customer's order status from an existing REST API. Which Azure AI Agents capability should you configure?

  1. A system-assigned managed identity alone
  2. A function tool that calls the API Correct answer.
  3. A second model deployment
  4. A larger thread

Explanation

Function tools let an agent invoke external code or APIs during a run. The agent emits a call request, your code executes it, and the result is returned to the run.

Question 42Develop multi-agent solutions in Azure

Which statement accurately describes the role of an agent's instructions in a multi-agent solution?

  1. They determine the pricing tier of the resource
  2. They specify the network topology of the deployment
  3. They encrypt the thread contents at rest
  4. They define the agent's persona, scope, and behavioral constraints Correct answer.

Explanation

Instructions are the system-level guidance that shapes how an agent responds, including its role, boundaries, and output expectations. They do not control infrastructure or security settings.

Question 43Develop multi-agent solutions in Azure

In a connected-agents design, what is the primary purpose of giving each specialist agent a narrow, well-defined responsibility?

  1. To improve reliability and make each agent easier to evaluate Correct answer.
  2. To reduce the number of model deployments required
  3. To eliminate the need for an orchestrator
  4. To avoid using threads

Explanation

Narrow responsibilities reduce ambiguity in instructions and tool selection, which makes agent behavior more predictable and testable. Orchestration and state management remain necessary regardless.

Question 44Develop multi-agent solutions in Azure

Which Azure resource provides the identity and access boundary under which an Azure AI Agents project operates?

  1. The client's local development machine
  2. The model weights file
  3. The Azure AI Foundry project and its connected resource Correct answer.
  4. The agent's instruction text

Explanation

The project and its connected Azure resources define the identity, permissions, and data boundary for agents created within it. Local machines, model files, and instructions do not establish that boundary.

Question 45Develop multi-agent solutions in Azure

What does an agent run return when it completes successfully?

  1. A copy of the model deployment key
  2. A run status plus the messages or tool calls produced Correct answer.
  3. A billing invoice
  4. A new thread identifier only

Explanation

A run reports its status and the outputs produced during execution, which may include messages and tool call requests. It does not create a new thread or expose credentials.

Question 46Develop multi-agent solutions in Azure

An orchestrator agent must wait for three worker agents to finish before composing a final answer. Which pattern is being applied?

  1. Fan-out/fan-in Correct answer.
  2. Human-in-the-loop
  3. Sequential handoff
  4. Router

Explanation

Fan-out distributes subtasks to multiple workers in parallel, and fan-in aggregates their results before the orchestrator continues. Sequential handoff would pass control one agent at a time instead.

Question 47Develop multi-agent solutions in Azure

You need the orchestrator to receive structured output from a specialist agent so it can branch on a category value. Which approach is most reliable?

  1. Shorten the thread by deleting earlier messages
  2. Increase the model temperature
  3. Ask the agent to answer in free-form prose
  4. Instruct the agent to return JSON and validate it against a schema Correct answer.

Explanation

Constraining output to a defined schema and validating it makes downstream branching deterministic. Free-form prose and higher temperature both increase variability.

Question 48Develop multi-agent solutions in Azure

A worker agent repeatedly calls the same function tool with identical arguments and makes no progress. Which control is most appropriate to add?

  1. A higher model temperature
  2. A maximum iteration or tool-call limit in the orchestration loop Correct answer.
  3. A longer system prompt
  4. A second thread per run

Explanation

Bounding the number of iterations or tool calls prevents runaway loops and lets the orchestrator fail gracefully. Prompt length, thread count, and temperature do not cap repetition.

Question 49Develop multi-agent solutions in Azure

Which mechanism lets an orchestrator pause a workflow until a human approves a proposed refund?

  1. A retry policy on the model call
  2. A scheduled thread cleanup job
  3. A human-in-the-loop checkpoint that suspends and later resumes the run Correct answer.
  4. A vector search over policy documents

Explanation

Human-in-the-loop checkpoints intentionally pause execution and require an external approval before proceeding. Retries, retrieval, and cleanup do not introduce an approval gate.

Question 50Develop multi-agent solutions in Azure

Two agents must share context about the same customer conversation. What is the recommended approach in Azure AI Agents?

  1. Have both agents operate on the same thread or pass summarized context explicitly Correct answer.
  2. Store context only in the model deployment name
  3. Create a separate thread for each agent with no shared data
  4. Rely on the agents guessing from prior knowledge

Explanation

Sharing a thread or explicitly passing summarized context ensures both agents reason over the same facts. Isolated threads or implicit assumptions produce inconsistent answers.