Generative AI Enterprise
Posts
Google's latest AI agents best practices

Google's latest AI agents best practices

Plus, Meta Llama 4, EU AI Act playbook, and more.

Lewis Walker
April 06, 2025

WELCOME, EXECUTIVES AND PROFESSIONALS.

Those delivering gen AI agents will know it's relatively straightforward to transition from idea to proof of concept, but realizing high-quality results in production is more complex. Google's latest best practices help address this challenge.

Since the previous edition, we've reviewed hundreds of the latest agentic and gen AI best practices, case studies, and innovation insights. Here’s the top 1%...

In today’s edition:

Google: Operationalizing AI agents.
Meta releases Llama 4 models.
PwC: Harnessing the EU AI Act.
2025 TCS Digital Twindex: Manufacturing.
Transformation and technology in the news.
Career opportunities & events.

Read time: 4 minutes.

BEST PRACTICE INSIGHT & CASE STUDIES

Google: Operationalizing AI agents

Image source: Google

Brief: Google’s 76-page “Agents Companion” paper builds on its original “Agents” paper, exploring the operationalization of gen AI agents. It covers AgentOps, evaluation, Agentic RAG, multi-agent systems, real-world case studies, and more.

Breakdown:

In operationalizing agents, Google emphasizes the importance of metrics within AgentOps to help build, monitor, and compare agent revisions.
For agent evaluation, it focuses on assessing agent capabilities, evaluating trajectory and tool use, and evaluating the final response.
AI is evolving towards multi-agent systems, specialized agents working together to achieve complex goals. The paper outlines design patterns.
Agentic RAG architecture is highlighted: autonomous retrieval agents that actively refine their search based on iterative reasoning.
Google showcases how specialized agents collaborate to power in-car conversational AI, illustrating real-world multi-agent systems in action.

Why it’s important: Gen AI agents mark a leap beyond standalone LLMs, enabling dynamic problem-solving and interaction. This “102” guide builds on core concepts, offering in-depth exploration of agent evaluation methods and practical applications to help enterprises operationalize results in production.

INNOVATION INSIGHT

Meta releases Llama 4 models

Image source: Meta

Brief: Meta released Llama 4 Scout and Maverick, open-weight, natively multimodal models with expanded context length. Its first mixture-of-experts (MoE) architecture models, they offer a blend of top performance and low-cost.

Breakdown:

Llama 4 Scout is the smallest model: 109B parameters, 17B active across 16 experts. It runs on a single GPU and has a 10M+ token context window.
Llama 4 Maverick has 400B parameters, 17B active across 128 experts. It requires more GPUs but runs as fast as Scout due to active parameters.
Maverick ranks second on LMArena, just below the Gemini 2.5 Pro Experimental, a reasoning model. Maverick is a non-reasoning model.
Maverick offers a leading mix of performance and cost ($0.19-$0.49 per 1M input & output tokens, 3:1 blended), as shown in the graph above.
Zuckerberg revealed there is more to come in the next month: Llama 4 Reasoning and Llama 4 Behemoth, featuring over 2 trillion parameters.

Why it’s important: Meta’s high-performance, low-cost Llama 4 models are a notable upgrade for enterprises, with Meta capturing 16% of the market. The release also introduces a new level of intelligence to consumers, with 700 million active Meta AI users across platforms like Instagram, WhatsApp, and Messenger.

AI GOVERNANCE

PwC: Harnessing the EU AI Act

Image source: PwC

Brief: PwC published an 11-page report outlining the EU AI Act requirements based on an organization’s role and AI system risk potential, with a focus on compliance through the implementation of effective governance.

Breakdown:

Establish AI standards covering the AI Act requirements and implement AI-specific controls to ensure traceability across the AI lifecycle.
Build an AI inventory with a meta model and registration process to enhance transparency of AI systems.
Apply a risk-based approach throughout the AI lifecycle, identifying, assessing, and mitigating risks, with clear entry gates for new use cases.
Define AI accountability across the lifecycle and create roles as needed to ensure responsibility and quality at each checkpoint.
Raise AI awareness within the organization so all employees can use AI responsibly and roles within AI governance are effectively executed.

Why it’s important: The EU AI Act, which applies to any company doing business in the EU or offering AI to EU customers, banned high-risk AI systems from February 2025. The full Act applies to all AI systems from August 2026. Non-compliance can incur fines up to EUR 40 million or 7% of annual turnover.

INDUSTRY INSIGHT & CASE STUDIES

2025 TCS Digital Twindex: Manufacturing

Image source: Tata Consulting Services

Brief: Tata Consultancy Services’ Digital Twindex, featuring executives from TCS, Jaguar Land Rover, and Siemens, explores how digital twins infused with gen AI, physical AI, and agentic AI help manufacturers anticipate and adapt to change.

Breakdown:

Gen AI is transforming design, planning, and simulation, while physical AI closes the loop between insight and real-world action.
Agentic AI enables systems to reason and adapt. Quantum computing is set to supercharge simulation across energy, logistics, and materials.
Digital twins are driving real-time decision-making, predictive maintenance, workflow redesign, factory-floor intelligence, and more.
Manufacturing is shifting to flexible, modular setups, supported by gen AI and AI micro-factories in select industry segments.
Airbus cut time and cost using digital twins in assembly. BMW leveraged intelligent systems to predict and prevent production issues.

Why it’s important: Digital twins were once on the fringes, but with today’s AI, they’ve become more strategic. Evolving from static simulations to dynamic, intelligent systems, they now fuse data, context, and computation to power real-time decision-making and connect physical operations with digital intelligence.

PwC shared a 14-slide report on data management, showing how gen AI helps enable privacy, security, governance, architecture, and quality.

BCG released an 88-page paper on AI’s potential in Indian agriculture and healthcare, including principles to build an adaptable AI strategy.

Deloitte finds that insurers aren’t fully ready for gen AI at scale, but with focus on resources, responsibility, and returns, POCs can scale to production.

Adobe shared a report on how gen AI imrpoves B2B marketing, plus 25 slides on success with clients like Prudential, Dentsu, and Red Hat.

Accenture released research on AI-powered ad platforms, highlighting the opportunity at stake with gen AI and how advertisers can harness it.

Stanford explored Human-AI collaboration for knowledge work, using M&A deals to show how agents can support research and decision-making.

Canva published its state of AI & marketing report, 47-slides of insights from 2,400 leaders, highlighting enterprise investments and successes.

AI 2027, a 70-page scenario analysis, predicts superhuman AI will surpass the Industrial Revolution's impact in the next decade.

Microsoft announced Actions, Memory, Deep Research, Pages, evolving Copilot into a smarter, more personal AI companion.

Amazon AGI Labs unveiled Nova Act, an AI agent system that controls web browsers to perform tasks, plus a developer SDK for creating agents.

OpenAI's ChatGPT reached 20M paid users, pushing OpenAI’s annual revenue to $5B. This follows a $40B funding round and GPT 4o’s image generation virality.

Intel is reportedly forming a joint venture with chipmaking rival TSMC to run the U.S. semiconductor manufacturing facilities.

AWS expands agent support by adopting Anthropic’s Model Context Protocol (MCP), standardizing AI model access to tools and data at run-time.

Microsoft made its agent framework generally available, an extension of Azure AI Foundry’s Semantic Kernel, designed to simplify multi-agent orchestration.

OpenAI plans to release its first open-weights model since GPT-2, according to Sam Altman, with pre-launch events to help developers get involved early.

MiniMax launched Speech-02, a new model with ultra-realistic text-to-speech and voice cloning across more than 30 languages.