Modern LLM-based agents are specifically designed to avoid requiring every possible execution path to be predetermined. The uploaded course material therefore correctly identifies "Requiring every execution path to be predefined" as the behavior that is NOT characteristic of an agent.
OpenAI defines agents as systems capable of independently accomplishing workflows using an LLM to manage workflow execution and make decisions. An agent can determine when a workflow is complete, correct its actions after receiving observations, and dynamically select tools according to the current state.
This differs fundamentally from conventional deterministic automation in which developers encode every branch and execution path beforehand.
Agents commonly pursue objectives across multiple reasoning-and-action cycles. They can invoke external APIs, databases, search systems, or other tools; inspect the resulting observations; and choose subsequent actions. A typical agent loop continues until an exit condition is reached rather than following one permanently fixed sequence.
Predetermined rules may still be used for safety, permissions, and guardrails, but the complete path toward the goal does not need to be pre-scripted.
Therefore, D is the correct answer.
Study Guide reference/topic: Introduction to AI Agents - autonomy, agent loops, observations, dynamic tool use, multi-step goal execution, and deterministic workflows.

The distinction concerns ownership of the conversation and orchestration flow , not synchronous versus asynchronous execution. In the OpenAI Agents SDK's Manager pattern-also called agents-as-tools-a central manager remains the active user-facing agent. It invokes specialist agents as tools, receives their outputs, synthesizes them, and retains responsibility for the final response. The specialist supports the manager without taking ownership of the conversation.
A Handoff works differently. When the current agent hands the task to another agent, the selected specialist becomes the active agent and takes over the conversation for the remainder of that portion of the run.
OpenAI's official Agents SDK documentation explicitly describes the Manager pattern as retaining control and Handoffs as transferring control to a specialized agent.
This distinction allows architects to select centralized orchestration when a single agent must aggregate results or apply common controls, and decentralized handoffs when specialists should directly own particular interactions. Therefore, D states the relationship correctly and matches the uploaded question's answer key.
Study Guide reference/topic: OpenAI Responses API and Agents SDK - multi-agent orchestration, Agents as Tools/Manager pattern, and Handoffs.

For application-defined function tools, an LLM does not inherently execute the external operation itself.
Instead, the model produces a structured tool-call request identifying the selected function and supplying arguments. The application or agent runtime then interprets that request, applies appropriate validation or authorization, invokes the corresponding implementation, and returns the result to the model for subsequent reasoning.
The OpenAI Responses API defines function calls as custom tools supplied by the developer that enable the model to request execution of application code using typed arguments. The OpenAI Agents SDK makes the separation explicit: a FunctionTool contains the tool name, description, parameter JSON schema, and an on_invoke_tool implementation that actually executes when the runtime processes the model-generated arguments.
This boundary is critical for security. The model proposes an action; controlled application/runtime logic performs the action. The model is therefore not automatically granted direct database, filesystem, network, or operating-system privileges merely because a tool has been described to it.
Options A, B, and D incorrectly remove this enforcement boundary. Consequently, C is the correct architectural description and matches the supplied question source.
Study Guide reference/topic: OpenAI Responses API and Agents SDK - function calling, structured tool calls, application execution, validation, schemas, and security boundaries.

MCP standardizes how external systems expose capabilities to an AI application, but the model does not need to reason about the transport or deployment location of each capability. Once an MCP server's tools are discovered and incorporated into an agent's available tool set, they are represented to the model as callable tools with names, descriptions, and input schemas. Locally implemented function tools are presented through essentially the same model-facing tool abstraction. OpenAI's Agents SDK documentation explicitly states that tools obtained from configured MCP servers are added to the agent's list of available tools, alongside ordinary tools. Therefore, from the LLM's perspective, both are selected and invoked through the tool-calling mechanism rather than through separate network-specific interfaces.
Authentication, network connectivity, server lifecycle, authorization, and actual execution remain responsibilities of the application/MCP infrastructure. They are deliberately abstracted away from the model.
Therefore, option B precisely captures the architectural consistency described in the course question.
Study Guide reference/topic: Model Context Protocol (MCP) Fundamentals - MCP tools, tool discovery, agent tool abstraction, and client-server integration.