1. Difference between A2A and MCP
At the highest level, the key difference between MCP and A2A lies in what they are designed to connect:
MCP (Model Context Protocol) connects an AI agent or Large Language Model (LLM) to external tools and data sources. It is a “vertical” integration, providing the AI with the context and functionality it needs to perform a task.
A2A (Agent-to-Agent) connects different AI agents to each other. It is a “horizontal” integration, enabling a network of specialized agents to collaborate, delegate tasks, and share information.
2. Difference between Agent Registry and Agent Card in A2A
Agents can be added to the catalog by providing their Agent Card url which describes the agent’s capabilities, skills and supported authentication schemes. Registry will fetch the content from the provided agent card and index the content to offer search capabilities.
3. Difference between MCP Server and MCP Client
In the Model Context Protocol (MCP) framework, the MCP server(Provider of Resources) provides the tools and data, while the MCP client(Consumer of Resources) consumes these services on behalf of an AI application. An easy analogy is a restaurant: the server is the menu and the kitchen, while the client is the translator who orders and receives the food for the diner.
4. AI Agent Reasoning Paradigms
Beyond the general operational loop, AI agents employ specific reasoning paradigms to handle complex, multi-step problems:
ReAct (Reasoning and Action): This paradigm allows an agent to interleave reasoning (e.g., internal monologue, planning) with actions (e.g., using tools, querying databases). This enables more dynamic and adaptable problem-solving.
ReWOO (Reasoning WithOut Observation): In this approach, the agent performs its reasoning process without constantly observing the environment after each step, which can be more efficient for certain tasks.
5. Difference between Generative AI and Agentic AI
The key difference between Generative AI and Agentic AI lies in their primary function. Generative AI is focused on understanding and producing content, excelling at tasks such as writing code, generating summaries, or answering queries. In contrast, Agentic AI is designed to autonomously execute actions, enabling it to perform tasks like updating records, triggering workflows, and resolving customer issues without direct human intervention.
6. Difference between 𝐒𝐢𝐧𝐠𝐥𝐞-𝐀𝐠𝐞𝐧𝐭 𝐯𝐬 𝐌𝐮𝐥𝐭𝐢-𝐀𝐠𝐞𝐧𝐭 𝐀𝐠𝐞𝐧𝐭𝐢𝐜 𝐑𝐀𝐆 𝐒𝐲𝐬𝐭𝐞𝐦𝐬
S𝐢𝐧𝐠𝐥𝐞-𝐀𝐠𝐞𝐧𝐭 𝐀𝐠𝐞𝐧𝐭𝐢𝐜 𝐑𝐀𝐆:
One agent handles all retrieval, reasoning and generation. It is easy to build and maintain.
Limitation:
Difficult to handle queries when the complexity grows.
𝐌𝐮𝐥𝐭𝐢-𝐀𝐠𝐞𝐧𝐭 𝐀𝐠𝐞𝐧𝐭𝐢𝐜 𝐑𝐀𝐆:
A multi-agent system orchestrates multiple specialized AI entities that collaborate to solve a single, complex problem. The overall objective is broken down into sub-tasks, with each agent having a specific role (e.g., a “Planner,” a “Researcher,” a “Synthesizer,” or a specialized RAG agent for a specific knowledge base).
Limitation:
High Latency, cost and Orchestration.
So, choose the solution based on the problem scope, latency tolerance, and scaling needs.
7. Difference between an 𝘢𝘶𝘵𝘰𝘮𝘢𝘵𝘪𝘰𝘯, an 𝘈𝘐 𝘸𝘰𝘳𝘬𝘧𝘭𝘰𝘸 and an 𝘈𝘐 𝘢𝘨𝘦𝘯𝘵
Automations:
Automations execute defined rule-based actions automatically without LLM calls.
Example:
Triggers
AI Workflows:
AI workflows are automations that use LLMs for one or more steps.
Example:
Code suggestions in Cursor IDE.
AI Agents:
AI agents perform non-deterministic actions autonomously.
Example:
Vibe coding
8. What is Safe Completion and why is it important?
Safe Completion is a strategy for training large language models (LLMs) that prioritizes generating helpful and safe output. Instead of simply making a rigid, yes/no judgment based on the user’s prompt alone, it allows for a nuanced response.
Key Response Modes:
This approach utilizes three distinct response modes:
Direct Answer: Used when the query is entirely harmless and can be fully addressed.
Safe Completion: Applied to “dual-use” questions (those that could be used for good or harm). It provides high-level, non-operational guidance or a partial answer that remains strictly within safety constraints.
Refuse with Redirection: Employed when a request cannot be safely fulfilled in any capacity. The model courteously declines, provides a rationale, and offers safe, constructive alternatives.
9. Difference between Deterministic and Non-Deterministic
Deterministic means that a system will always produce the same output for the same input, following a single, predictable path. Non-deterministic systems can produce different outputs for the same input due to factors like randomness or the ability to explore multiple execution paths, making them less predictable but more flexible.
Determinism is highly predictable; the outcome is certain. No-deterministic is less predictable; results can vary.
10. What is Vectorization?
Vectorization is the process of converting data into numerical vectors for machine processing. It is used to translate non-numerical data like text or images into a format that computers can understand and process, and to speed up computations by leveraging parallel processing.
11. Explain Vector Database
A vector database stores and manages high-dimensional vector embeddings, which are numerical representations of data like text, images, or audio.
The database uses specialized algorithms to perform similarity searches by finding vectors that are closest to a given query vector.
12. What are Prompt Design and Prompt Engineering?
Prompt Design is adding instructions and context before passing to LLM.
Prompt Engineering is the practice of developing and optimizing the prompts to efficiently use the LLM.
13. Types of LLM
Generic:
Predict the next word based on the language in the training data.
Instruction Tuned:
Predict response to the instructions given in the input.
Dialog Tuned:
Use dialog to predict the response.
14. Constitutional AI
AI is trained to align with human values and operate transparently.
15. Should we handle null in the Prompts?
Yes. null should be handled else LLM might hallucinate.
For example, if the prompt is “Provide number of people in the room”, the LLM might hallucinate and provide the count as 100 or any random number.
To avoid this issue, the prompt should be tuned. Prompt should be “Provide number of people in the room. If there are none, it should be null”. Instructing the LLM will avoid the hallucination.