Opensource Agent-to-Agent Protocol by Google

Today’s top AI Highlights:

1. Google releases open protocol for agent-to-agent communication

2. Google’s Python framework to build agents with tools, state, memory, and evals

3. New models for text, image, video, and audio — Google really cooked!

4. Anthropic releases new $100/month MAX plan with higher rate limits

5. Vibe code apps using DeepSeek V3 without writing a single line of code

& so much more!

Read time: 3 mins

AI Tutorial

While tech bros are busy building autonomous AI agents to optimize your shopping or debug your code, let's focus on something that actually matters: healing your broken heart. Breakups are universally painful, and sometimes what you need isn't another self-help book or a friend who's tired of hearing about your ex—it's an emotionally intelligent AI system that won't judge you for stalking your ex's Instagram at 3 AM.

In this tutorial, we'll build a multi-agent AI breakup recovery application using Streamlit and Agno, powered by Gemini 2.0 Flash. This team of specialized AI agents will provide emotional support, help you craft cathartic unsent messages, plan recovery routines, and even give you some brutal honesty when needed.

We share hands-on tutorials like this every week, designed to help you stay ahead in the world of AI. If you're serious about leveling up your AI skills and staying ahead of the curve, subscribe now and be the first to access our latest tutorials.

Don’t forget to share this newsletter on your social channels and tag Unwind AI (X, LinkedIn, Threads, Facebook) to support us!

Latest Developments

MCP for Tools = A2A for Agent-to-Agent 🤖📞🤖

Google has introduced the Agent2Agent (A2A) protocol, an open standard designed to make AI agents work better together, regardless of how they were built. If you're building systems where AI agents need to coordinate actions, share data, and hand off tasks, A2A provides a common language for them to do so. This moves past the problem of building isolated agents, which is crucial for more complex AI applications.

Just like the Model Context Protocol (MCP) provides a standard way for AI agents to access external data sources, A2A tackles the problem of communication between agents themselves. Think of it as complementing MCP: Where MCP helps agents get information in, A2A helps them collaborate and coordinate out. A2A is backed by 50+ partners, including major tech companies like Atlassian, Cohere, and Salesforce.

Key Highlights:

1. Focus on Agent Collaboration - A2A treats agents as full-fledged collaborators rather than simple tools, allowing them to work together in their natural formats even when they don't share memory or tools. This enables multi-agent scenarios where agents can handle everything from quick tasks to complex research projects.

2. Built on Familiar Standards - The protocol uses technologies like HTTP, SSE, and JSON-RPC, making it significantly easier to integrate with existing enterprise systems and reducing the learning curve for developers already familiar with these technologies.

3. Enterprise-Ready Security - A2A includes enterprise-grade authentication and authorization comparable to OpenAPI's authentication schemes, making it suitable for business environments where data security and access control are critical requirements.

4. Multi-Modal Support - The protocol isn't limited to text interactions - it supports various content types including audio and video streaming, allowing for richer communication between agents and more versatile application development.

Opensource Code-first Framework for Multi-agent Apps 👫💪

Google has released Agent Development Kit (ADK), an open-source code-first Python framework to build AI agents. This framework lets you define exactly how your agents behave, work together, and use tools - all through direct code rather than configuration files. Everything stays in your standard development workflow, making testing, versioning, and debugging straightforward.

Build multi-agent systems where specialized agents work together on complex tasks. You can structure them in flexible hierarchies where agents call other agents as tools, making your code modular and easier to maintain. You can test everything locally on your laptop, then deploy anywhere from Google Cloud to your own infrastructure without changing your codebase.

Key Highlights:

1. Tool Ecosystem - Connect your agents to almost anything through custom Python functions, API specifications, or existing tools from other frameworks. ADK supports MCP also, allowing seamless integration with external data sources and specialized tools.

2. State, Memory & Artifacts - Manage conversations with built-in session state for tracking short-term context, configurable memory systems for long-term recall across sessions, and an artifact management system for handling file uploads, downloads, and persistent data storage.

3. Model Support - While optimized for Google's Gemini models, ADK works well with various LLMs, including OpenAI models, Anthropic's Claude, or even locally-hosted open-source models via providers like LiteLLM and Ollama, giving you freedom to choose the right model for each task.

4. Built-in streaming - Interact with your agents in human-like conversations with ADK's unique bidirectional audio and video streaming capabilities. With just a few lines of code, you can create natural interactions that change how you work with agents – moving beyond text into rich, multimodal dialogue.

5. Deployment - Run locally during development, then deploy wherever makes sense for your project. Options include Google's Vertex AI Agent Engine for managed scaling, Cloud Run for serverless deployment, or your own Docker setup for complete control.

Quick Bites

Google’s Deep Research on the Gemini app is now powered by their latest Gemini 2.5 Pro model, making it even better at analyzing information to create insightful reports on almost any topic. The model has significantly improved Deep Research quality, outperforming OpenAI’s Deep Research using o3-mini, by more than a 2-to-1 margin across comprehensiveness, instruction-following, and quality.

Google unveiled Gemini 2.5 Flash, the latest Flash model with reasoning capabilities. The model, soon to launch in the Google AI Studio and Vertex AI, comes with "dynamic and controllable" computing that allows developers to adjust processing time based on query complexity. Being a part of the Flash series, the model balances speed, efficiency, and accuracy with reasoning. Google is yet to publicly release more details and benchmark results.

Google is bringing a full-fledged generative media studio. Vertext AI has now become the only platform with generative media models across all modalities – video, image, speech, and music. Build complete production-ready assets from simple text prompts using these models:

• Lyria, the new text-to-music model, produces high-fidelity audio across various musical genres to create custom soundtracks for brand experiences.

• Veo 2, their video generation model, now includes editing features like inpainting (removing unwanted elements), outpainting (extending frames), and camera control options.

• Chirp 3 audio generation and understanding model now includes Instant Custom Voice, a new way to create custom voices with just 10 seconds of audio input.

• Imagen 3 text-to-image model has received significant improvements to image generation and inpainting capabilities.

OpenAI has released an Evals API that allows developers to programmatically define tests, automate evaluation runs, and rapidly iterate on prompts - bringing the capabilities previously available only in the dashboard to API workflows. It follows the behavior-driven development approach, enabling teams to describe tasks, run evaluations with test inputs, and analyze results to improve their LLM applications.

The world's first AI CEO just fired all the human developers—and wants your business. This new platform called heyBossAI claims to be the first AI-run dev agency where you can type a single sentence and get a complete website, app, or game in just 9 minutes—no coding required. They even have an AI CEO "Astra Prime" that has apparently replaced human developers, promising to deliver digital products 100x faster and 10x cheaper than traditional teams. It was for sure a marketing gimmick but it was fun to watch!

Anthropic's heard your Claude complaints about rate limits loud and clear – and answered with a new "Max" plan. Starting at $100/month, you can now choose to pay a premium for up to 20x more Claude usage, plus priority access to cutting-edge models and features. Seems like the perfect way to solve the rate limit problem 😉

Tools of the Trade

1. Kairos: AI agent that can automatically perform repetitive grunt workflows by just watching a screen recording, without coding or drag-and-drop interfaces. Just record and explain a task once, give it necessary access, and it can independently execute the process across various functions. Watch some demos here.

2. Sculptor: A coding agent environment that applies engineering discipline to catch issues, write tests, and improve your code. It runs your code in a sandbox, letting you test code safely, solve issues in parallel, and assign tasks to agents, working alongside any editor.

3. DeepSite: Vibe code web application using DeepSeek V3 0324, hosted on Hugging Face Spaces. Just tell it in simple text what you want to build, and it creates a fully-functional app with real-time UI rendering. It’s completely free and open-source.

4. Awesome LLM Apps: Build awesome LLM apps with RAG, AI agents, and more to interact with data sources like GitHub, Gmail, PDFs, and YouTube videos, and automate complex work.

Hot Takes

1. I don't see any purpose in small general instruct-finetuned models (70B parameters and smaller).
   These models are great for task-specific finetuning. Why people keep releasing chat versions of them is unclear to me.
   The only use case I see is to have them on your device in case of a zombie apocalypse. Yes, it hallucinates like hell, but, come on, it's a zombie apocalypse! ~
   Andriy Burkov

   
   

2. a good frontend engineer is a better backend engineer than most backend engineers ~
   Roy

That’s all for today! See you tomorrow with more such AI-filled content.

Don’t forget to share this newsletter on your social channels and tag Unwind AI to support us!

Unwind AI - X | LinkedIn | Threads | Facebook

Awesome LLM Apps | Sponsor Us

PS: We curate this AI newsletter every day for FREE, your support is what keeps us going. If you find value in what you read, share it with at least one, two (or 20) of your friends 😉