Enterprise AI Agents and Multi-Agentic Systems with Google Cloud: From Concept to Production

Introduction: Agentic Solutions from Concepts to Scale

In the rapidly evolving landscape of artificial intelligence, two concepts are revolutionizing business processes & workflows today: AI agents and multi-agentic systems. These technologies are helping companies move beyond simple chatbots and robotic process automation (RPA) to full-blown rewiring of the enterprise. In this blog series intended for strategic AI executives and hands-on developers alike, we go deep into these technologies and explain how you go about deploying Agents & achieve business outcomes for your enterprise using Google Cloud technologies [1].

This multi-part series is an outcome of our experiences from real world Agentic solutions & projects for multiple Google Cloud customers and derives from the work done by Tredence [11] and the Google Partner Engineering team.

Blog Series Overview

Our goal is to present the emerging Agentic AI patterns in practice and answer common questions for AI leaders and practitioners. We will first explain what AI Agents are and take you through a tour of the complete lifecycle of designing, building, deploying, and operating AI Agents and multi-agentic solutions. 

In this series we will cover the following topics:

1. Introduction to Agents and Multi-agent Systems - the Why and What (This blog)
2. Agent Development with ADK and Vertex AI - Marketing Agent example (with code)
3. Agents with Tools - MCP tools integration with Agents in Vertex AI (with code)
4. Building Multi-Agentic Systems (MAS): Using A2A protocol with agents across multiple SDKs & frameworks (with code)
5. The Google Agentic Stack: Agentspace, Vertex AI Agent Engine etc.
6. Evaluating Agentic AI and Generative outcomes (with code)
7. AI metrics, operations and lifecycle management at enterprise scale

We will use Python programming language and Google Cloud technology stack with Vertex AI [1], Agentspace, Agent Development Kit (ADK) [6], and Agentic protocols such as A2A (Agent-to-Agent) and MCP (Model Context Protocol) for our explanations. However, the ideas & patterns we present are broadly applicable to other languages, cloud stacks, and open-source frameworks.

After a gentle introduction in the first blog, this series will get technically deep in the later chapters, but we promise it will be fun & useful. So let’s go!

AI Agents - The Evolution from 2023 to Today!

Google simply defines AI agents as software systems that use AI to pursue goals and complete tasks on behalf of users. Here are some ways to characterize AI agents in real world:

• AI Agents show reasoning, planning, and memory and have a level of autonomy to make decisions, learn, and adapt [2].
• AI agents are intelligent, autonomous entities that can go well beyond predefined rules and are capable of executing multi-step tasks and handling multi-turn conversations.
• AI Agents utilize tools [6] to access the internet and other data in the company for performing complex tasks and collaborating with each other to help achieve goals defined for them. 

Here’s a good easy way to depict AI Agents [10]:

Diagram: AI Agents with Reasoning, Planning, Memory and Tools to connect with the external world.

The concept of Agents is not new. For example, in the excellent book, “Artificial Intelligence: A Modern Approach”, Peter Norvig & Stuart Russel [5] explained agents well before ChatGPT came by! 

However, over the past three years, the definition of an "AI agent" has undergone a dramatic and rapid evolution - from a long-standing academic concept to a tangible and powerful reality that is being implemented across consumer, business and engineering domains. Before 2023, an “Agent” was supposed to have some rule-based behavior, some limited autonomy and was explicitly programmed to perform a specific task. With the rise of Large Language Models (LLMs) such as OpenAI’s GPT’s, Google’s Gemini family of models, Llama, Claude, and others, agents can now reason, plan, and autonomously execute complex multi-step tasks to achieve specific goals. In the last 6-12 months, new protocols such as MCP and A2A have emerged to standardize how AI agents can talk to each other and to other systems. And as Andrej Kaparthy says, “The future involves designing software explicitly for LLM-based agents” which is happening now. All this has been less of an “evolution” and more of a “revolution”.

Diagram: Evolution of AI Agents from LLM’s over the last few years.

As specialists working in the field of Generative AI and Agentic AI implementations, we often get asked a lot of questions about how AI agents can be built and used safely. Here are some of the most common questions and their answers.

How are AI Agents different from traditional software or Generative AI? What makes AI Agents so capable?

AI Agents have some unique characteristics that make them different from traditional software: 

1. AI Agents use Large Language Models (LLMs) - LLM’s essentially act as the brain and give agents the power of logical reasoning & planning to make recommendations and decisions.
2. AI agents are goal-driven & have a higher degree of autonomy - Unlike traditional software that follows predefined rules, AI agents can plan their execution paths to achieve pre-defined goals for them. They learn, adapt, and operate with a fair degree of independence. 
3. AI agents perceive their environment and interact with it. This includes the ability to retry tasks that failed, and also interaction with humans if needed. Typically, agents have built-in prompts or system instructions that help them perform specific functions such as sending an email.
4. AI Agents help humans make decisions and they take actions on the user’s behalf, orchestrating complex workflows that were previously manual and time-consuming.
   

Generative AI is good at creating new content, it doesn’t necessarily have a focused “goal” and it doesn’t need to take actions. On the other hand, AI Agents have these capabilities built on top of LLMs - typically act on behalf of their users by using tools to access APIs and other external systems. 

An AI agent may have a user interface (UI) and may look like a standard chatbot similar to Google’s Gemini or OpenAI ChatGPT. However, that’s not necessary. We are developing multiple agents right now that are doing complex work in the background with no visible UI.

As a result of context, reasoning, and the ability to execute multi-step complex workflows, AI Agents are much more powerful & capable than traditional process automation methods such as RPA. 

What are Multi-Agentic Systems?

Multi-agentic systems (MAS), as the name suggests, involve multiple AI agents interacting and collaborating to solve a problem or achieve a shared goal. As AI agents handle more and more complex tasks, it makes sense to break a monolithic agent into smaller composable agents that can work together to perform the task. Turns out, just like with us humans, Agents are much more effective when they collaborate and work in teams! 

Each agent in the system typically has specialized capabilities & goals, and by working together, they  tackle challenges that would be too complex for a single agent. This collective intelligence allows for more robust, scalable, and sophisticated solutions.

Example MAS: Agentic AI for Digital Marketing Automation

Here’s an example of a MAS solution we implemented earlier this year: Lauren, a product marketing manager at a large CPG company, runs marketing campaigns. It used to take her 2-3 months to plan and execute a campaign earlier. Now she uses a multi-agent "Agentic Marketing Assistant" solution. This solution has specialized agents for campaign analysis, content creation, refinement, and testing. It allows her to quickly ideate and create targeted marketing content. This solution makes her life easy and enables her to launch multiple targeted campaigns faster - she now launches a campaign in 2-4 weeks. For her company, this has meant significantly reduced costs and increase in the market share.

MAS Architectural Patterns

Agent development frameworks like ADK, Langgraph, etc. facilitate multi-agent collaboration by composing them together in multiple MAS patterns [5] and [6]. For example, a MAS may have a coordinator agent which orchestrates with other specialized agents. Or the agents can be invoked in a sequential order. The right pattern depends on the use case, how agents interact with each other and tools, their autonomy, and balance between human and AI collaboration. We will dive deeper into these patterns and architecture options in our future blog.

Below are common patterns with ADK and examples for these patterns.

Diagram: MAS patterns supported by ADK.

What are people really using AI Agents for?

Most of us are already using AI agents, often without even realizing it! 

AI Agents for Consumers

AI agents are a part of the mobile apps, websites, and other online services that we use every day.  

For example, our personal assistants and smart devices like Apple Siri, Google Assistant, smartwatches and other similar devices & services are essentially  agents. For example, the Garmin watch monitors activity, sleep patterns, and heart rate, etc. for a person and tells them how they should adjust their training routines.

Agents are also integrated into many of our favorite apps. In your day-to-day life, agents like Google’s Gemini are in Gmail, Google Search, and Google Workspace to help you draft emails, summarize documents, and provide complex answers. Companies like Uber use AI agents behind the scenes to optimize the rider experience and manage operations. Other companies and apps like Best Buy, IHG Hotels & Resorts, Home Depot, Etsy and Mercedes Benz are also building AI-powered agents to help with customer service and online shopping experiences.

You can now also use AI Agents for generating new content. A kid can easily create a professional looking Mother’s Day card using AI agent powered tools such as Canva or Imagen.

AI Agents for Software Developers

Perhaps the biggest and most direct impact AI Agents are having today is in the space of software development and coding. With Coding Agents, you can often generate large amounts of useful code by just writing instructions (prompts) in English. 

At Tredence, our teams have embraced AI-powered coding Assistants to improve their productivity by 20-60%, depending on the specific type of engineering work they’re doing. They are using AI-powered IDE’s such as Cursor or getting help from a coding agent such as GitHub Copilot and Gemini Code Assist.

Google’s Gemini Code Assist, Code Assist Agent Mode and Gemini CLI  are AI-powered coding assistants from Google. They support code completion, explanation, code chat, use agentic chat as a pair programmer, bug fixing, and code generation using natural language prompts. We will cover them in detail in later parts of this blog series.

At Tredence, we have developed Data Engineering Agents and Data Science Agents that compliment Google’s BigQuery Data Agents. These offerings help with data ingestion, data quality, pipeline management, ML model creation, testing, and other aspects of data engineering, machine learning and full-stack Generative AI application lifecycle.

Agents are not just helping write code, but they’re completely changing the software development lifecycle by automating CI/CD, writing and executing test cases, updating JIRA tickets, going through logs for debugging, and monitoring production applications.If your teams are not effectively using these tools, you are missing out on a big part of developer productivity and need to modernize your teams and tools urgently.

How are enterprises using AI Agents?

As a specialist AI and Data services company, Tredence is involved in many Agentic AI solution implementations. Here are some key examples of use cases we have helped companies build and deploy with AI Agents & multi-agentic systems in the last 12 months:

• Intelligent Data Analysis: With Agentic Decision Intelligence, companies are deploying agentic systems to sift through vast amounts of data, identify patterns, generate insights, and even make predictions, supporting strategic decision-making.
• Supply Chain Optimization: Agents can analyze real-time data on inventory, demand, and logistics to optimize routes, manage stock levels, and minimize disruptions.
• Intelligent Workflow Management: We have built AI agents for orchestrating complex business processes, from data collection and analysis to task assignment and approval routing.
• Better Customer Engagement: AI agents powering personalized customer interactions through sophisticated virtual assistants, providing tailored support, answering queries, and guiding customers through processes.
• Proactive Issue Resolution: By monitoring systems and data, agents are detecting anomalies or potential problems and initiate corrective actions before they escalate.
• Fraud Detection and Security: We are building solutions to monitor network traffic, track financial transactions and user behavior patterns to identify suspicious activities and flag potential security breaches.
• Software Lifecycle Acceleration: AI Agents are now helping developers generate code, debug their software, and run the software delivery lifecycle much faster.

For more examples of how leading enterprises are building AI agents, refer to this resource from Google, which highlights 601 use cases: Real-World Generative AI Use Cases [9].

What benefits does Agentic AI provide to companies and their employees? Are there any risks involved?

Agentic AI Risks

As with any new technology that is fast evolving, Agentic AI comes with its own risks and challenges. Outside of the ethical and societal risks that are broadly discussed elsewhere, here are the most pressing AI execution risks that we have seen and some guidelines on how organizations can begin to build guardrails around them.

• Risk: Over-promising and Under-delivering: Executives often do not fully grasp the complexity and evolving nature of Agentic AI. Their AI leaders tend to promise unrealistic ROI, setting up projects for failure. Many existing solutions and approaches are pitched under the “agentic” guise, leading to underachievement of results.

Guardrail: Anchor projects in well-defined business outcomes & identify measurable success criteria. Establish iterative pilots with measurable ROI before scaling.

• Risk: Integration Complexity: Integration between Agentic AI solutions and existing enterprise systems is the key to success of Agentic and Generative AI applications [10]. And successful integration requires a deep understanding of a business's processes.
  Organizations tend to underestimate the complexity involved in properly designing this integration across the data, API and the application layers. MCP servers and A2A protocol have made decent progress, but these technologies are in their early stages.

Guardrail: Treat integration as a first-class design problem. Invest in partners with integration expertise, develop blueprints, test agents in sandboxed environments, and involve process owners early.

• Lack of Generative AI and Agentic AI skillsets and tools: As pointed out well in the survey outlined in the MIT study [10], enterprises attempting to build Generative AI solutions on their own are much more likely to fail than those who purchase AI tools from specialized Cloud vendors and work with partners who are specialized in AI implementations.

Guardrail: Carefully choose your Agentic AI technology stack and develop implementation partnerships with AI expertise. Ensure that inhouse talent is developed over time and available to ensure ongoing success of your Agentic AI projects.

• Safety & Security Risks: Prompt injection, memory poisoning, loss of data privacy through incorrect data access by models, etc. are all real risks.

Guardrail: Be aware of the data security & privacy risks involved in agent implementations. Involve your CISO organization early to establish clear guidelines and secure platforms to build on. Ensure that your AI system takes care of PII protection, blocks NSFW content, and defends against prompt injection, for example. Put all implementations through expert security reviews.

• Lack of explainability and governance structure: Companies, especially in the regulated industries such as Financials and Healthcare, are still figuring out the organizational governance as well as technical solutions for understanding & controlling Agentic solutions.

Guardrail: You can mitigate this risk through proper instrumentation of your Agentic AI applications at platform level to capture key events and calls, capture of key metrics with cross-layer tracing, and proper protocols for reporting and response. 

Active research is underway, and solutions are being launched every week to address key issues such as explainability, i.e., understanding and tracking agents make decisions, controlling which agents can access which data, how to keep humans in charge, and how to get regulatory compliance with agentic solutions

Some of these risks and issues are still not fully resolved and are likely to remain so in the near future. However, that doesn’t mean that you should wait. The cost of waiting is too high: it can lead to digital obsolescence and loss of leadership in your business. 

In the future articles of this series, we will go through how you can evaluate agents and put safety guardrails around your implementations to control and mitigate these risks.

How Do I Get Started with Building AI Agents?

Based on your business requirement and complexity of the use case, you can choose between a wide variety of platforms and tools to build your agentic solutions. For example, to build a search-based application connecting with your internal documents for your employees, you could choose Google Agentspace. For custom agents and complex workflows, developers often prefer to use frameworks such as LangGraph and ADK. 

The Agent Development Kit (ADK) is a powerful, open-source Python framework designed to help you create, test, and deploy production-grade AI agents and multi-agent systems [6]. ADK is particularly important for enterprises because it's built from the ground up for multi-agent systems, where specialized agents work together to tackle complex problems. 

ADK is a flexible and modular framework that supports building a variety of agent types, allowing you to choose the right one for your specific use case.

• LLM Agents: These agents use a large language model (LLM) as their "brain" to understand natural language, reason, plan, and make dynamic decisions. They're ideal for flexible, language-centric tasks where the execution path isn't strictly predefined.
• Workflow Agents: These are designed for orchestrating the execution flow of other agents in a predefined, deterministic manner. ADK provides three main types of workflow agents:
  • SequentialAgent: Executes tasks one after another in a specific order.
  • ParallelAgent: Allows multiple agents to work on tasks simultaneously.
  • LoopAgent: Enables agents to repeat tasks until a specific condition is met.
• Multi-Agent Systems (MAS): ADK's core strength is its support for MAS, which involves composing multiple specialized agents that collaborate to achieve a larger goal. This allows you to break down complex problems into manageable tasks, with different agents handling specific functions.

Critical Considerations for Agentic Planning & Design

Before you start building, it's essential to define the purpose and capabilities of your AI agent and go through a battle-tested list of items to consider. To ensure we are building a high-quality solution that serves a real business purpose, at Tredence we use the following checklist to guide our planning and design process -

1. Strategic Questions: Defining the Agent's Purpose

• Who are the intended users of the system, and what functions will they use this solution for? In other words, what is the agent's primary objective?
• What are the success criteria for the agentic solution being implemented? Define both the business metrics and non-functional criteria in measurable terms.
• What inputs does the agent expect, and what outputs does it return?
• What tools (APIs, Model Context Protocol (MCP) servers) or data sources will the agent need to effectively achieve its goals?
• What framework, platform, or tools will you use to build the agent (e.g., ADK, LangChain, etc.)?

2. Designing for a Production-Ready Agent

To build a secure, reliable, and maintainable agent, you must address several design and development aspects.

Security and Credentials

• Will the agent process personal or sensitive data? How will it ensure the safety of this data? For example, redaction, access control, etc.
• Are you managing API keys and other credentials securely?
• How will model calls and interactions be logged? Can the calls be traced from user interactions to the underlying systems and LLMs?
• How will you test the agent for adversarial inputs or prompt injections?
• Will the agent clearly disclose that it's AI-powered and communicate its limitations?

Testing and Evaluation

• What is your evaluation methodology, and what tools will you use to test the agent's performance?
• How will you measure the agent's trajectory and tool use, and the quality of its final response?
• How will the agent handle invalid inputs or failures, and can it "self-heal" by reading error messages and correcting its next action?
•  Is there a human fallback or escalation process in place if the agent can't respond or if a human is needed?
• Have the prompts and system instructions been reviewed for clarity, tone, and safety?
• Are the outputs helpful, brand-aligned, policy compliant, and appropriate for varied user inputs?
• How will the agentic solution collect end user feedback on quality etc., and how will you use this information (e.g., RLHF, tuning of prompts, UX changes, etc.).

Context Engineering

• How will you manage the information the AI model sees, including writing context (saving it outside the context window), selecting relevant context to pull in, compressing it to retain only necessary tokens, and isolating it to help an agent perform a specific task?

3.    Building and Deploying

• What services and runtime will be used to execute the agent?
• How will you manage the agent's development lifecycle, including building, testing, deploying, and monitoring?
• What interfaces will users, apps, and other agents use to interact with the agent (e.g., simple API endpoint, Agent-to-Agent (A2A) protocol)?
• What is the expected average response time for a typical query, and are there retry mechanisms and rate-limit handling in place?
• How will you measure the agent’s effectiveness compared to your existing process?

Conclusion

In this blog, we explained the Agentic AI concepts and gave some real-world use cases for multi-agentic systems. In the later parts of this blog series,  we will deep dive into the best practices and hands-on approaches for each of the above critical questions. The solutions we develop will demonstrate how you can drive productivity, reduce execution risks, and significantly improve your customers’ experience within your enterprise systems & software. 

Our explanations and examples are based on real world implementations that we are involved in, but remember that this is a fast evolving field and things are changing every day. So refer to the latest documentation and release notes as you work through these examples.

Transform your enterprise with Google Cloud and Tredence - Learn more and schedule a quick consultation-Powering Advanced AI & Data Transformation with Google Cloud and Agentifying Business Processes

References

[1] Google Cloud Vertex AI Platform, https://cloud.google.com/vertex-ai

[2] LLM Agents and Planners, Memory, etc. ADK Documentation, https://google.github.io/adk-docs/agents/llm-agents/#planner and https://google.github.io/adk-docs/sessions/memory/  Understanding the planning of LLM agents: A survey by Xu Huang et al, https://arxiv.org/pdf/2402.02716

[3] Large Language Model Agents - Berkeley MOOC, Fall 2024 https://rdi.berkeley.edu/llm-agents/f24

[4] AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversation, https://arxiv.org/abs/2308.08155 

[5] Artificial Intelligence: A Modern Approach, Peter Norvig and Stuart Russell, 4th Edition 2021  

[6] Agent Development Kit Resources: Common Multi-Agent Patterns using ADK Primitives, https://google.github.io/adk-docs/agents/multi-agents/#common-multi-agent-patterns-using-adk-primitives

ADK Tools, https://google.github.io/adk-docs/tools/

[7] Multi-agent Systems with LangGraph, https://langchain-ai.github.io/langgraph/concepts/multi_agent/#multi-agent-architectures

[8] ReAct: Synergizing Reasoning and Acting in Language Models, Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, Yuan Cao, https://arxiv.org/abs/2210.03629

[9] Real-World Generative AI Use Cases, Google Cloud, https://cloud.google.com/transform/101-real-world-generative-ai-use-cases-from-industry-leaders

[10] The GenAI Divide: State of AI in Business 2025, MIT NANDA Report - 95% generative AI pilots at companies are failing,  (behind paywall/restricted access) https://fortune.com/2025/08/18/mit-report-95-percent-generative-ai-pilots-at-companies-failing-cfo/

[11] Agentifying Business Processes, Tredence, https://www.tredence.com/services/agentic-ai#service

[12] Gemini Code Assist, https://codeassist.google/

AUTHOR - FOLLOW
Unmesh Kulkarni
SVP, AI, Tredence Inc.

AUTHOR - FOLLOW
Maruti C
Global Partner Engineering Leader, Data & AI, Google Cloud