In today’s newsletter:

• Little Book of Deep Learning

• MCP vs A2A Clearly explained

• RaBitQ – Efficient vector compression with high recall and ranking preservation

Reading time: 3 minutes.

The Little Book of Deep Learning!

If you're looking for a clear and concise guide to deep learning, this might be the best one out there right now.

Written by François Fleuret, a professor of computer science at the University of Geneva. This resource walks through the full deep learning stack.

Covers core topics like mathematical foundations, efficient computation, model architectures, training methods, and generative models

👉 Find the book here (PDF version)

MCP vs A2A

Why MCP vs A2A Matters

As agentic AI rapidly evolves, two protocols have become foundational for building scalable and interoperable systems: MCP (Model Context Protocol) and A2A (Agent-to-Agent). Understanding the unique roles they play and how they work together is essential for anyone designing, deploying, or scaling agentic AI.

Agentic LLMs: Protocols at the Core

Modern agentic systems rely on protocols to enable intelligent, autonomous behavior.

• MCP is designed to let a single agent securely access tools, APIs, and data sources in a standardized way.

• A2A enables multiple agents to communicate, coordinate, and delegate tasks to each other, even if they come from different vendors or frameworks.

This protocol-driven approach ensures agents can not only act independently but also collaborate to solve complex, real-world problems.

A2A Core Concepts

A2A follows a client-server model. A main agent (client) delegates tasks to specialized agents (servers) that expose their capabilities through structured HTTP endpoints.

Key Components:

• Agent Card: A JSON file that lists the agent’s capabilities, endpoint URL, and authentication method.

• Task: The unit of work. Each task has a lifecycle: submitted, working, completed, or failed.

• Message: Structured communication between agents, exchanged within the context of a task.

• Part: Content blocks within messages (text, file, or structured data).

• Artifact: The final result of a completed task, returned to the client.

How A2A Works: Step by Step

1. Discovery
   The client reads Agent Cards to identify which agents can assist with the task.

2. Task Initiation
   It sends a structured task request to the appropriate agent’s HTTP endpoint.

3. Processing
   The remote agent performs the work, asks for more information, or sends progress updates.

4. Multi-Turn Conversations
   Agents can exchange messages to refine task inputs and clarify requirements.

5. Status Updates
   Long-running tasks can stream updates using server-sent events or push mechanisms.

6. Completion
   The agent returns the result as an artifact. The client compiles outputs into a final response.

A2A uses an “opaque agent” model: agents expose what they can do, not how they do it. This allows private logic and proprietary workflows to remain hidden while still participating in shared systems.

A2A Communication

• Feature identification
  Identify agents based on published capabilities in their Agent Cards.

• Workflow Coordination
  Track lifecycle state and ownership of each task.

• Collaboration
  Enable structured, multi-message exchanges to complete complex tasks.

• UX Customization
  Adapt content formats depending on what the receiving agent can display (e.g., text, video).

MCP vs A2A: Solving Different Layers

Agentic applications benefit from using both protocols:

A useful way to think about it:

• MCP provides vertical integration: Connecting an application (and its AI model) deeply with the specific tools and data it needs.

• A2A provides horizontal integration: Connecting different, independent agents across various systems.

Imagine MCP gives an individual agent the knowledge and tools it needs to do its job well. Then, A2A provides the way for these well-equipped agents to collaborate as a team.

These protocols don’t overlap, they complement each other. You can use both together right now.

RaBitQ – New Approach to Efficient Vector Search

Scaling vector databases often means trading off between memory usage, accuracy, and search speed.

Traditional indexing methods like HNSW, PQ, and OPQ each compromise in different ways. You either lose recall, increase memory usage, or slow down performance.

Milvus introduces RaBitQ (Rank-aware Binary Tree Quantization), a new method that reduces memory usage by 72 percent, maintains around 95 percent recall, and delivers up to 4x faster full-text search. It also improves support for multilingual text analyzers.

By preserving the relative ranking of nearest neighbors during binary quantization, RaBitQ enables faster and more accurate hybrid and multi-modal search across both text and vector data.

👉 Read more about out RaBitQ

That’s a Wrap

That’s all for today. Thank you for reading today’s edition. See you in the next issue with more AI Engineering insights.

PS: We curate this AI Engineering content for free, and your support means everything. If you find value in what you read, consider sharing it with a friend or two.

Your feedback is valuable: If there’s a topic you’re stuck on or curious about, reply to this email. We’re building this for you, and your feedback helps shape what we send.

WORK WITH US

Looking to promote your company, product, or service to 120K+ AI developers? Get in touch today by replying to this email.