Google’s Agent-To-Agent Interoperability Protocol

Also, OpenAI publishes EU policy blueprint recommendations

Author

Jack Lajoie
April 09, 2025 • read time ~ 9 minutes

⚡️ Headlines

🤖 AI

IBM releases a new mainframe built for the age of AI - IBM launches the z16 mainframe, designed to handle AI workloads with integrated on-chip AI accelerators and enhanced security for enterprise environments.

Trump says the future of AI is powered by coal - Donald Trump promotes coal as essential to the future of AI infrastructure, sparking backlash from environmental and tech communities.

Amazon unveils a new AI voice model, Nova Sonic - Amazon introduces Nova Sonic, a generative AI voice model capable of natural-sounding speech, claiming superior speed and cost-efficiency compared to competitors.

Deep Research powered by Gemini 2.5 Pro (experimental) now available - Google's Gemini Advanced subscribers can now utilize Deep Research with the 2.5 Pro model, offering enhanced analytical reasoning and information synthesis capabilities.

Mira Murati's AI startup gains prominent ex-OpenAI advisers - Thinking Machines Lab, led by former OpenAI CTO Mira Murati, has added ex-OpenAI researchers Bob McGrew and Alec Radford as advisers, aiming to develop customizable and capable AI systems.

Nvidia Faces Dilemma After Chinese Firms Rush to Order $16 Billion in New AI Chips - Chinese tech companies have reportedly placed orders totaling $16 billion for Nvidia's AI chips, presenting the company with strategic challenges amid geopolitical tensions.

Ai2 and Google Cloud commit $20M to advance AI-powered research for the Cancer AI Alliance - Ai2 and Google Cloud each invest $10 million into the Cancer AI Alliance to develop AI models aimed at accelerating cancer research while ensuring data privacy.

You can now give Google's AI video model camera directions - Google's Veo 2 model now allows users to apply cinematic techniques and edit videos through text prompts, enhancing video content creation capabilities.

Introducing Cogito Preview - DeepCogito releases Cogito v1, a series of open-license large language models trained using Iterated Distillation and Amplification, outperforming existing models in various benchmarks.

China: Quantum computer does world-first fine-tuning of billion parameter AI model - Chinese scientists achieve a milestone by using the Origin Wukong quantum computer to fine-tune a billion-parameter AI model, demonstrating quantum computing's potential in AI training.

🦾 Emerging Tech

Justice Department disbands cryptocurrency enforcement team - The U.S. Department of Justice is disbanding its National Cryptocurrency Enforcement Team, shifting focus to prosecuting individuals who use digital assets for crimes such as terrorism and human trafficking.

🤳 Social Media

Instagram's chief outlines strategies for creators to increase reach - Adam Mosseri emphasizes the importance of posting to the main feed, leveraging direct messages for content sharing, and optimizing for enhanced SEO capabilities to boost visibility on Instagram.

Snap introduces generative AI ad format focusing on user engagement - Snapchat launches sponsored AI lenses, allowing users to interact with branded, AI-generated visuals, aiming for a more organic and engaging advertising experience.

🔬 Research

Researchers achieve one-minute video generation using Test-Time Training - A novel method employing Test-Time Training layers enables pre-trained Diffusion Transformers to generate coherent one-minute videos from text storyboards, outperforming previous models in temporal consistency and motion smoothness.

🔌 Plug-Into-This

Google DeepMind has introduced the Agent Communication Language (ACL) protocol, a proposed standard to enable seamless, structured interaction between autonomous AI agents. Drawing inspiration from decades of multi-agent systems research, the initiative seeks to establish a shared “lingua franca” for agentic collaboration across heterogeneous environments and vendors.

  • ACL messages are designed to be machine-readable, verifiable, and grounded in specific semantics to prevent ambiguity and manipulation.

  • The protocol supports nested goal structures, temporal constraints, and epistemic states, enabling rich negotiation, delegation, and collaboration workflows between agents.

  • Implementations are being piloted in DeepMind's open-source Melting Pot framework and tested in domains such as document processing, supply chain planning, and robotic control.

  • Google invites contributions from academia and industry to shape the specification into an open standard for ecosystem-wide adoption.

  • ACL builds on historical efforts like KQML and FIPA ACL but incorporates contemporary alignment safeguards and cryptographic verification layers.

  • Think of this like giving AI agents a standardized “email + calendar + contracts” system to coordinate with each other—across brands, platforms, or even tasks.

🧩 Google's move signals a shift from siloed LLM-based agents toward interoperable ecosystems, echoing early internet standardization phases. The big question now: will open protocols outpace proprietary agent stacks in defining the coordination substrate of multi-agent AI?

OpenAI has published an “Economic Blueprint” tailored to the European Union, outlining strategies to amplify regional productivity, address workforce transitions, and foster public-sector adoption of AI. The document is part of OpenAI’s broader engagement with EU institutions ahead of the AI Act’s rollout.

  • The blueprint identifies six high-impact economic domains, including education, health, manufacturing, and public services, as ripe for LLM-driven productivity gains.

  • It advocates for "sovereign fine-tuning"—allowing European institutions to adapt foundation models using localized datasets and values.

  • Proposals include LLM-powered copilots for bureaucratic workflows, SMEs, and policy development, aiming to streamline institutional efficiency.

  • OpenAI emphasizes partnerships with local developers and regulators to ensure culturally and linguistically aligned AI deployments.

  • The report underscores the importance of reskilling and social safety nets as automation pressures intensify across job sectors.

  • In plain terms: OpenAI is making a pitch to the EU—“let us help modernize your economy with our tech, but on your terms.”

🇪🇺 This blueprint illustrates OpenAI’s ambition to shape—not just adapt to—Europe’s regulatory and economic AI agenda. It positions the company not merely as a model provider, but as a policy co-author.

Stanford HAI’s annual AI Index Report 2024 presents a data-rich snapshot of the global AI landscape, highlighting an unprecedented rise in private AI investment, expanding foundation model dominance, and intensifying regulatory discourse. This article highlights the most interesting graphs.

  • Global private AI investment in 2023 hit $92B, with the U.S. capturing over 60% of it, and model training costs reaching hundreds of millions per run.

  • Benchmarks like MMLU and HellaSwag are saturating—newer models show marginal gains, prompting calls for more dynamic, real-world evaluations.

  • Open-source models saw rapid growth in capabilities, but are still largely trailing frontier closed models in aggregate performance.

  • The report flags a dramatic rise in AI policy activity: 127 AI-related legislative proposals across 33 countries were recorded in 2023.

  • Use of synthetic data in training pipelines nearly doubled, reflecting a shift in how models are being scaled and adapted.

  • In simple terms: AI’s moving fast, benchmarks are breaking, money’s flooding in, and governments are scrambling to keep up.

📊 The report highlights an inflection point—AI innovation is accelerating faster than our ability to measure or regulate it. As benchmark fatigue sets in, the focus may shift to longitudinal, use-case-driven evaluations and governance frameworks that can keep pace.

 🆕 Updates

📽️ Daily Demo

🗣️ Discourse