AI Advanced: Agent Frameworks & ADK Practice Questions

SequentialAgent is built for fixed, ordered pipelines. It runs each sub-agent in the order listed and passes outputs downstream. An LlmAgent that reasons about call order at runtime uses an LLM to decide what to call, which adds unneeded non-determinism and cost to a fixed pipeline. Running the summarizer and translator simultaneously with ParallelAgent would fail because the translation depends on the summary output. LoopAgent's repeat-until-termination behavior is meant for iterative or retry patterns, not a linear one-pass pipeline.

Whether an LLM uses a tool depends heavily on how the tool is described in the system prompt and tool schemas. If the AgentTool's description is vague, or the system prompt does not steer the model toward it for this kind of task, the LLM may answer directly instead of delegating. The claim that AgentTool is incompatible with LlmAgent is wrong; working with an LlmAgent is AgentTool's main use case. The expired-session explanation sounds plausible but is wrong, because AgentTool invocations create their own execution context and do not depend on an existing persistent session. The LoopAgent-wrapper idea is wrong because LoopAgent is for iterative execution, not for forcing a tool call.

The statement that tasks run strictly one at a time in defined order is correct. Because the sequential process runs tasks one after another, three truly parallel database searches are not natively supported. The developer would need the hierarchical process or async task tooling for that. The statement that each task's output is passed as context to the next is also correct. In the sequential process, earlier task outputs are automatically available as context to later tasks, and that is how the SynthesisAgent would receive the researchers' findings. The load-balancing claim is wrong: the sequential process does not assign tasks by agent workload. It just follows task order. The manager_llm claim is wrong: a manager LLM is required only for the hierarchical process. The shared mutable global memory claim is wrong: CrewAI has no shared global memory that tasks write to concurrently. Tasks communicate through structured task outputs and context passing.

Task success rate measures whether the agent produced the correct final output, which directly covers dimension (1). Trajectory efficiency or step-count evaluation measures whether the agent used the minimal, expected number of tool calls to reach the answer, which directly covers dimension (2). BLEU score measures text similarity to reference answers and does not suit evaluating whether an agent's answer is functionally correct. Token cost per task measures money or compute spent, not whether tool usage was logically minimal. Average end-to-end latency measures speed, not redundant tool calls. Precision/Recall on entity extraction is a task-specific NLP metric, not a general agent evaluation metric for dimension (1) in this scenario.

LlmAgent is correct because it is backed by an LLM that performs chain-of-thought reasoning to decide which tools to invoke and when, making it suitable for open-ended, dynamic tasks. SequentialAgent simply chains sub-agents in a fixed order and has no LLM-driven reasoning of its own. LoopAgent repeatedly runs sub-agents until a termination condition is met but does not itself use an LLM for decision-making. ParallelAgent runs sub-agents concurrently in parallel but also does not have its own LLM reasoning layer.