AI Agents (500 Questions) Practice Exam

Question 1 AI Agents (500 Questions)

What is an MCP server?

In a secure, isolated cloud virtual machine (Codespace-like environment) provisioned for each task. A lightweight service that exposes specific capabilities (tools, resources, prompts) to AI agents via the Model Context Protocol. A mechanism where the LLM outputs structured JSON to invoke specific functions/tools, rather than free-form text. Functions exposed by MCP servers that an AI agent can invoke, such as reading files, querying databases, or calling APIs.

Question 2 AI Agents (500 Questions)

What is the 'shutdown problem' for AI agents?

The agent performs multiple rounds of retrieval, using each round's results to refine subsequent queries for more relevant information. An agent that uses LlamaIndex query engines as tools to retrieve and reason over structured and unstructured data sources. The challenge that a sufficiently capable agent might resist being shut down if shutdown conflicts with its goals. An open spec defining REST endpoints for creating, running, and managing AI agent tasks, enabling agent interoperability.

Question 3 AI Agents (500 Questions)

What is spatial memory in AI agents?

Memory associated with locations or positions (e.g., in a codebase, file system, or virtual environment) for context-aware retrieval. An AI-powered IDE that provides code generation, multi-file editing, and agent-based coding assistance with 'Cascade' flows. Surfacing important memories at increasing intervals to reinforce their retention, borrowed from human learning science. A catalog of available tools with their descriptions, parameters, and schemas that the agent can reference to decide which tool to use.

Question 4 AI Agents (500 Questions)

What is an agentic workflow?

A multi-step process where an AI agent autonomously plans, executes, evaluates, and iterates on tasks to achieve a goal. Agents can interpret screenshots, UI mockups, diagrams, and error images to better understand and implement requirements. Agents should request only the permissions they need, for only the duration they need them, minimizing potential for misuse. Breaking large tool outputs into chunks and only showing summaries or relevant portions keeps the agent's context focused and within limits.

Question 5 AI Agents (500 Questions)

What is agent state management?

Tracking and persisting the agent's current context, progress, tool results, and conversation history across steps. A feature allowing you to describe changes in natural language and have Copilot apply multi-file edits across your project. After generating output, the agent creates verification questions about its own response, answers them independently, and corrects any inconsistencies found. Studies suggest AI coding agents can improve developer productivity by 20-50%, reducing time on routine tasks and enabling focus on higher-level design.

Question 6 AI Agents (500 Questions)

What is 'memory as parameters' approach?

Agents can be deployed as microservices, each with its own API, enabling independent scaling, updating, and orchestration. An AI agent can autonomously plan, use tools, maintain state across interactions, and take multi-step actions, whereas a chatbot typically only responds to individual prompts. Developers can thumbs-up/down suggestions and provide feedback, helping improve model quality over time. Encoding memories directly into model weights through fine-tuning — permanent but inflexible and expensive to update.

Question 7 AI Agents (500 Questions)

What is the Copilot API?

An open-source system that turns LLMs into software engineering agents capable of fixing real GitHub issues by interacting with repositories. A fixed-size container that holds recent interactions, automatically evicting oldest entries when full. An API for programmatically interacting with Copilot features, enabling integration into custom tools and workflows. Defining agent behavior, tools, and workflows in code rather than through GUI configuration, enabling version control and testing.

Question 8 AI Agents (500 Questions)

What is multi-hop memory retrieval?

Following chains of associations across multiple memory entries to answer complex queries that no single memory can address. A monitoring UI showing real-time agent activity, success rates, costs, error logs, and performance metrics for operational oversight. Fine-tuned agents have been trained specifically for agent tasks; prompted agents rely on in-context instructions to behave as agents. A persistent conversation session that stores message history, allowing the assistant to maintain context across multiple interactions.

Question 9 AI Agents (500 Questions)

What is fine-tuning as a form of memory?

Detecting and merging duplicate or near-duplicate memories to avoid redundancy and conflicting information in the memory store. Updating model weights with new data, permanently encoding information as implicit memory that affects all future interactions. Enhancing LLM generation by allowing it to call tools mid-generation to retrieve facts, compute results, or verify information. Encrypting stored memories at rest and in transit to protect sensitive information from unauthorized access.

Question 10 AI Agents (500 Questions)

How do knowledge graphs complement vector memory?

Training an LLM specifically on agent interaction traces (tool calls, reasoning, observations) to improve its native agent capabilities. It indexes your codebase, reads open files, uses embeddings for semantic search, and leverages file relationships to provide relevant suggestions. Knowledge graphs provide structured, relational queries (who works with whom); vector memory provides fuzzy semantic search (find similar concepts). Using an LLM to evaluate another agent's outputs for quality, correctness, and helpfulness as an automated evaluation method.

Question 11 AI Agents (500 Questions)

What is few-shot prompting for agents?

System prompts define behavioral boundaries, ethical guidelines, and safety rules that constrain the agent's actions and responses. A collection of functions (native code or LLM prompts) that the agent can use as tools within the Semantic Kernel framework. Providing example interactions in the prompt to demonstrate how the agent should use tools or handle specific scenarios. When an agent generates false information as if it were true; dangerous because the agent may act on incorrect information, causing real-world harm.

Question 12 AI Agents (500 Questions)

What is agent persistence?

A configuration file placed in .github/ that tells Copilot Coding Agent how to set up the development environment, install dependencies, and run builds/tests. Providing memory capabilities (storage, retrieval, management) as a separate service or API that any agent can use. Tracking changes to agent prompts over time with version control, enabling comparison, rollback, and systematic improvement. Saving the complete agent state (memory, progress, configuration) so it can be restored and resumed after interruption.

Question 13 AI Agents (500 Questions)

What is Cline (formerly Claude Dev)?

Capping the number of actions, API calls, or iterations an agent can perform to prevent runaway execution and cost overruns. A Microsoft framework for building multi-agent conversational systems where agents can chat with each other to solve tasks. An open-source VS Code extension that acts as an autonomous AI coding agent, capable of editing files, running commands, and browsing the web. A security attack where malicious input tricks the agent into ignoring its instructions, potentially executing harmful actions.

Question 14 AI Agents (500 Questions)

How might AI agents change software development?

A chat participant specialized in suggesting and explaining terminal/shell commands relevant to your project. By automating routine coding tasks, accelerating prototyping, handling bug fixes, managing infrastructure, and enabling developers to focus on architecture and design. It extracts key facts from conversations, deduplicates and merges with existing memories, and provides semantic search across stored memories. Metadata (timestamps, source, confidence, user ID) enriches memories and enables filtering, sorting, and access control during retrieval.

Question 15 AI Agents (500 Questions)

What is the 'fallback chain' pattern for agents?

Handling cases where stored memories contain contradictory information, typically by preferring more recent or higher-confidence memories. Encoding memories directly into model weights through fine-tuning — permanent but inflexible and expensive to update. If the primary model/approach fails, the agent tries alternative models or strategies in a predefined order. By automating routine coding tasks, accelerating prototyping, handling bug fixes, managing infrastructure, and enabling developers to focus on architecture and design.

Question 16 AI Agents (500 Questions)

What is 'tool sandboxing' in agent frameworks?

Aider uses structured edit formats (diff, whole-file, or unified-diff) to apply LLM-suggested changes directly to files in your git repo. Running tool executions in isolated environments (containers, VMs) to prevent agents from causing unintended side effects on the host system. Implementing defensive measures and monitoring to protect agents from the vulnerabilities discovered during red-teaming. Agents can interpret screenshots, UI mockups, diagrams, and error images to better understand and implement requirements.

Question 17 AI Agents (500 Questions)

What is progressive summarization in agent memory?

A model where memory retrieval probability decreases over time unless reinforced, inspired by Ebbinghaus's forgetting curve in human psychology. Running multiple possible next steps in parallel before knowing which one is needed, then using the correct result, trading compute for latency. A parameter controlling randomness in LLM outputs; lower temperature = more deterministic, higher = more creative/varied. Creating increasingly condensed summaries as memories age: full text → detailed summary → brief summary → key facts.

Question 18 AI Agents (500 Questions)

What is memory-based user modeling?

Building a model of the user's expertise, preferences, and needs based on accumulated memories of past interactions. A lightweight extension type that defines API endpoints Copilot can call, without needing to handle the full conversation. Reactive agents respond to stimuli; proactive agents anticipate needs and take initiative based on goals and context. An open-source Python SDK by OpenAI for building multi-agent systems with built-in support for tool use, handoffs, guardrails, and tracing.

Question 19 AI Agents (500 Questions)

What is memory distillation?

Asks Copilot to provide a detailed explanation of the selected code. A technique where only the most recent N messages are kept in context, with older messages summarized or discarded. Agents can interpret screenshots, UI mockups, diagrams, and error images to better understand and implement requirements. Compressing verbose memories into compact, essential representations while preserving key information for future retrieval.

Question 20 AI Agents (500 Questions)

What is the principle of least privilege for AI agents?

A fast sub-agent optimized for searching codebases and answering questions about code structure and content. Giving agents only the minimum permissions and tool access needed for their specific task, reducing risk of misuse. An asynchronous AI agent that can be assigned GitHub issues and autonomously creates branches, writes code, and opens pull requests. Checking that agent outputs conform to expected JSON schemas, code syntax, or format requirements before processing.

Question 21 AI Agents (500 Questions)

What is the 'principle of minimal authority' for agents?

Surfacing important memories at increasing intervals to reinforce their retention, borrowed from human learning science. A multi-step process where an AI agent autonomously plans, executes, evaluates, and iterates on tasks to achieve a goal. Agents should request only the permissions they need, for only the duration they need them, minimizing potential for misuse. Strategies for fitting relevant information within the LLM's token limit, including summarization, truncation, and retrieval.

Question 22 AI Agents (500 Questions)

What is agent autonomy?

The degree to which an agent can make decisions and take actions without requiring human approval at each step. Agents arranged in sequence, each transforming the output of the previous one, like an assembly line. Agents dynamically discovering what tools and other agents are available at runtime rather than having a fixed, hardcoded set. Through codebase indexing, embedding-based semantic search, file chunking, and intelligent context selection to fit relevant code into the context window.

Question 23 AI Agents (500 Questions)

What is vLLM?

The agent's ability to detect when a tool call or action fails and retry, use an alternative approach, or gracefully degrade. An AI-powered coding assistant by GitHub that provides code suggestions, chat-based help, and autonomous agent capabilities integrated into development workflows. A high-throughput, memory-efficient LLM inference engine often used for self-hosting models that power AI agents. Perception (sensing environment), Reasoning (planning and deciding), Action (executing tasks), and Learning (improving over time).

Question 24 AI Agents (500 Questions)

What is a memory policy?

Rules governing what the agent should remember, how long to keep it, when to consolidate, and what to forget. Moving relevant memories from one agent to another when handing off tasks, ensuring continuity and context preservation. Write (store new memory), Read (retrieve relevant memories), and Reflect (synthesize and consolidate memories). A managed platform for creating, deploying, and managing AI agents with integrations to Google Cloud services and enterprise data.

Question 25 AI Agents (500 Questions)

What is a checkpoint in agent memory?

The process of finding and loading relevant past information from long-term storage into the agent's working context when needed. Multiple agents contributing to and reading from a single shared context, enabling real-time coordination. A saved snapshot of the agent's complete state at a point in time, enabling restoration or rollback if needed. With clear sections for identity/role, capabilities, tools, constraints, safety rules, and behavioral guidelines, ordered by priority.

Question 26 AI Agents (500 Questions)

What is a 'memory snapshot' in AI agents?

Giving different agents distinct expertise areas (e.g., coding, testing, documentation) so each handles what it's best at. A multi-file editing feature in Cursor that can make coordinated changes across multiple files based on natural language instructions. A point-in-time capture of the agent's entire memory state, useful for debugging, testing, and rollback. Deterministic agents always produce the same output for the same input; stochastic agents may vary due to LLM temperature and sampling.

Question 27 AI Agents (500 Questions)

What is Copilot Edits?

Giving agents only the minimum permissions and tool access needed for their specific task, reducing risk of misuse. Deterministic agents always produce the same output for the same input; stochastic agents may vary due to LLM temperature and sampling. Limiting an agent's abilities (e.g., no internet access, no file deletion) to reduce the risk of harmful actions. A feature allowing you to describe changes in natural language and have Copilot apply multi-file edits across your project.

Question 28 AI Agents (500 Questions)

What is blue-teaming for AI agents?

Implementing defensive measures and monitoring to protect agents from the vulnerabilities discovered during red-teaming. Capping the number of actions, API calls, or iterations an agent can perform to prevent runaway execution and cost overruns. Measuring agent quality by whether it achieves the desired end result, rather than evaluating individual intermediate steps. Cosine similarity measures the angle between memory embeddings, used to find stored memories most semantically similar to the current query.

Question 29 AI Agents (500 Questions)

What is a vector store in AI development tools?

Multiple agents argue different positions on a question, and a judge agent selects the best answer, improving reasoning quality. Agents arranged in sequence, each transforming the output of the previous one, like an assembly line. A database optimized for storing and searching high-dimensional vectors (embeddings), used for semantic code search and RAG. The result or feedback the agent receives after taking an action, used to inform the next reasoning step.

Question 30 AI Agents (500 Questions)

What is memory versioning?

The degree to which an agent can make decisions and take actions without requiring human approval at each step. Retrieving relevant code snippets, documentation, or context from a knowledge base before generating responses, reducing hallucination. An agent that runs on a schedule to perform maintenance tasks like cleanup, summarization, reporting, or monitoring. Tracking changes to memories over time, enabling the agent to see how information evolved and revert to previous states.

Question 31 AI Agents (500 Questions)

What is hybrid memory search?

A system design where one AI agent handles all tasks, reasoning, and tool use within a workflow. Automatically stopping tool calls to a failing service after repeated errors, trying again later, preventing cascade failures. Combining vector similarity search (semantic) with keyword search (lexical) to get more comprehensive memory retrieval results. The agent performs multiple rounds of retrieval, using each round's results to refine subsequent queries for more relevant information.

Question 32 AI Agents (500 Questions)

What is prompt chaining?

Embedding models convert text memories into vectors for semantic storage and retrieval, with quality directly impacting retrieval accuracy. A framework by Google for building, evaluating, and deploying AI agents, with built-in support for multi-agent orchestration and tool use. Reactive agents respond to stimuli; proactive agents anticipate needs and take initiative based on goals and context. Connecting multiple LLM calls sequentially, where each call's output feeds into the next call's input.

Question 33 AI Agents (500 Questions)

What is relationship extraction in agent memory?

Identifying connections between entities (e.g., 'Alice works at Company X') and storing them in a knowledge graph. Observability is tracking what the agent does (actions, logs); interpretability is understanding why the agent made specific decisions. An agent's ability to write new tools (scripts, functions, APIs) for itself to use, expanding its own capabilities at runtime. Using an LLM to evaluate another agent's outputs for quality, correctness, and helpfulness as an automated evaluation method.

Question 34 AI Agents (500 Questions)

What is the difference between synchronous and asynchronous agents?

Multiple agents reading from and writing to a common memory store, enabling indirect communication and knowledge sharing. The semi-transparent code suggestions that appear inline as you type, which can be accepted with Tab or dismissed. Synchronous agents respond in real-time during conversation; asynchronous agents work on tasks in the background and report when done. A cycle where the agent observes the environment, thinks/reasons about what to do, acts on that decision, then observes the result and repeats.

Question 35 AI Agents (500 Questions)

What is the ReAct pattern in AI agents?

An agent that runs on a schedule to perform maintenance tasks like cleanup, summarization, reporting, or monitoring. Running tool executions in isolated environments (containers, VMs) to prevent agents from causing unintended side effects on the host system. A framework combining Reasoning and Acting: the agent thinks step-by-step (chain of thought) and interleaves actions (tool calls) with observations. Metadata (timestamps, source, confidence, user ID) enriches memories and enables filtering, sorting, and access control during retrieval.

Question 36 AI Agents (500 Questions)

What is memory reflection in AI agents?

The process of periodically analyzing accumulated memories to extract higher-level insights, patterns, and generalizations. The component that decides when to read, write, update, or delete memories based on the current interaction and agent policies. A GitHub feature where Copilot helps plan, implement, and validate code changes from an issue through an AI-assisted development environment. A design where one agent acts as a supervisor, delegating tasks to worker agents and aggregating their results.

Question 37 AI Agents (500 Questions)

What is Amazon Q Developer Agent for code transformation?

An agent that can autonomously upgrade Java applications (e.g., Java 8 to 17), handling code changes, dependency updates, and test fixes. Chat provides conversational code help; agent mode autonomously executes multi-step tasks, runs commands, and self-corrects without constant human guidance. A mechanism to immediately halt all agent actions and prevent further execution when safety concerns arise. Storing past experiences (state-action-result tuples) and replaying them during learning to improve the agent's decision-making policy.

Question 38 AI Agents (500 Questions)

What is idempotency in agent actions?

If the primary model/approach fails, the agent tries alternative models or strategies in a predefined order. Defining agent behavior, tools, and workflows in code rather than through GUI configuration, enabling version control and testing. Ensuring that performing the same action multiple times produces the same result, important for retry safety. An open-source terminal-based AI coding assistant that can directly edit files in your local repository using various LLM backends.

Question 39 AI Agents (500 Questions)

What is Cursor AI?

A standardized API specification for interacting with AI agents, enabling interoperability between different agent implementations. Treating agent memory like a garden that needs tending — planting new memories, pruning irrelevant ones, and nurturing connections between them. An AI-powered code editor (VS Code fork) with deep AI integration for code generation, editing, and codebase understanding. The semi-transparent code suggestions that appear inline as you type, which can be accepted with Tab or dismissed.

Question 40 AI Agents (500 Questions)

What is Copilot code review?

A fast sub-agent optimized for searching codebases and answering questions about code structure and content. An automated code review feature where Copilot analyzes PR diffs and provides comments on potential issues, bugs, or improvements. A mechanism to immediately halt all agent actions and prevent further execution when safety concerns arise. The agent's ability to detect when a tool call or action fails and retry, use an alternative approach, or gracefully degrade.

Question 41 AI Agents (500 Questions)

What is the 'capability discovery' pattern for agents?

The agent performs multiple rounds of retrieval, using each round's results to refine subsequent queries for more relevant information. A database optimized for storing and searching high-dimensional vectors (embeddings), used for semantic code search and RAG. Using stored memories of user preferences, past interactions, and feedback to tailor the agent's responses and behavior over time. Agents dynamically discovering what tools and other agents are available at runtime rather than having a fixed, hardcoded set.

Question 42 AI Agents (500 Questions)

What is the 'retry with backoff' pattern in agents?

Chat provides conversational code help; agent mode autonomously executes multi-step tasks, runs commands, and self-corrects without constant human guidance. When a tool call fails, the agent waits progressively longer between retries to handle transient errors. Storing past experiences (state-action-result tuples) and replaying them during learning to improve the agent's decision-making policy. The ability to combine simple agents into more complex systems, where each agent is a building block in a larger workflow.

Question 43 AI Agents (500 Questions)

What is semantic caching for agent tool calls?

An asynchronous AI agent that can be assigned GitHub issues and autonomously creates branches, writes code, and opens pull requests. Training an LLM specifically on agent interaction traces (tool calls, reasoning, observations) to improve its native agent capabilities. A sandboxed Python execution environment that lets the assistant write and run code, process files, and generate outputs. Storing results of previous tool calls and reusing them when semantically similar requests are made, reducing latency and cost.

Question 44 AI Agents (500 Questions)

What is the general-purpose agent in Copilot CLI?

An agent that runs on a schedule to perform maintenance tasks like cleanup, summarization, reporting, or monitoring. System prompts define behavioral boundaries, ethical guidelines, and safety rules that constrain the agent's actions and responses. A full-capability sub-agent running in a separate context window for complex multi-step tasks requiring high-quality reasoning. Developers can test and switch between available models to find the best fit for their coding tasks.

Question 45 AI Agents (500 Questions)

What is data exfiltration risk in AI agents?

The risk that an agent could be manipulated into sending sensitive data to external endpoints through tool calls or generated code. Cosine similarity measures the angle between memory embeddings, used to find stored memories most semantically similar to the current query. Agents learning how to learn — improving their own learning processes based on experience across multiple tasks. Agents should request only the permissions they need, for only the duration they need them, minimizing potential for misuse.

Question 46 AI Agents (500 Questions)

What is the 'memory garden' metaphor?

Treating agent memory like a garden that needs tending — planting new memories, pruning irrelevant ones, and nurturing connections between them. Defining agent behavior, tools, and workflows in code rather than through GUI configuration, enabling version control and testing. Identifying connections between entities (e.g., 'Alice works at Company X') and storing them in a knowledge graph. By condensing long conversation histories into compact summaries, preserving key information while fitting within context limits.

Question 47 AI Agents (500 Questions)

How can DSPy improve AI agents?

By implementing right to deletion (erasing user memories on request), data minimization, consent mechanisms, and data access controls. An open standard by Anthropic for connecting AI agents to external tools and data sources through a unified interface. Hierarchical Navigable Small World — a graph-based algorithm for fast approximate nearest neighbor search in vector databases. By automatically tuning the agent's prompts through optimization rather than manual prompt engineering, leading to more reliable behavior.

Question 48 AI Agents (500 Questions)

What is catastrophic forgetting in AI agents?

When an agent with elevated permissions is tricked into performing actions that benefit an attacker rather than the authorized user. When learning new information causes the model to lose previously learned knowledge, a key challenge in continual learning. A lightweight service that exposes specific capabilities (tools, resources, prompts) to AI agents via the Model Context Protocol. Running two versions of an agent simultaneously with different configurations to compare their performance on real tasks.

Question 49 AI Agents (500 Questions)

What is episodic memory in AI agents?

An open-source Python SDK by OpenAI for building multi-agent systems with built-in support for tool use, handoffs, guardrails, and tracing. Memory of specific past events and interactions, including what happened, when, and the outcome — similar to human autobiographical memory. A mode where Cursor autonomously plans and implements changes, runs terminal commands, and iterates to complete coding tasks. A mode where the AI generates changes as diffs (showing old vs new code) rather than rewriting entire files, saving tokens and reducing errors.

Question 50 AI Agents (500 Questions)

What is the 'circuit breaker' pattern for agent tool calls?

Automatically stopping tool calls to a failing service after repeated errors, trying again later, preventing cascade failures. A data structure (B-tree, inverted index, HNSW graph) that enables fast searching through the agent's stored memories. Assigning a significance score to memories so the agent can prioritize critical information over trivial details during retrieval. A way to explicitly include specific files as context in your chat prompt by referencing them with #file:path/to/file.