Google's New Agent Development Kit: A Business Perspective
- armstrongWebb
- 7 days ago
- 5 min read
AI Agents: Beyond the Basic Workflow
Let's cut through the tech jargon and get straight to what matters. Traditional software follows rigid, predictable paths—if X, then Y. You've been using this kind of technology for decades.
Click a button, get a result. Input data, receive output. It's consistent, but utterly incapable of handling the messy, nuanced reality of actual business problems.
AI agents represent something fundamentally different.
They don't just execute commands—they understand context, make judgment calls, and adapt on the fly. While your typical software application might break when facing an unexpected scenario, AI agents can reason through novel situations, pull in relevant tools, and actually figure things out.
The distinction is crucial: traditional workflows need explicit instructions for every possible path, while agents need objectives and capabilities. You're no longer micromanaging every digital step—you're delegating tasks to assistants that can 'think.'
Google ADK: What Is It, Really?
Google recently unveiled its Agent Development Kit (ADK) at Cloud NEXT 2025, and if you're running a business with complex problems to solve, you should be paying attention.
In essence, ADK is an open-source framework that dramatically simplifies building and deploying AI agents. It's Google's answer to a pressing question: how do we make sophisticated AI capabilities accessible to organisations without requiring a team of AI specialists?
The framework allows developers to create AI agents that can understand natural language, use various tools, reason through problems, and even collaborate with other agents. And it's designed to integrate with Google's ecosystem while remaining flexible enough to work with other platforms.
The ADK Advantage: Why This Matters for Your Business
Google isn't the first to release an agent development framework, but ADK brings distinct advantages that deserve your attention:
Simplicity with Sophistication: ADK makes complex agent systems feel more like standard software development. Your team can focus on business logic rather than wrestling with the underlying AI complexities.
Multi-Agent Architecture: Most frameworks build single agents. ADK is designed from the ground up for multiple specialised agents working in concert—mimicking how real teams tackle complex problems.
Enterprise-Ready Integration: While other frameworks might offer impressive demos, ADK is built for real business environments with direct pathways to production deployment and Google Cloud integration.
Model Flexibility: Your business isn't locked into using just Google's models. ADK supports various LLMs (Large Language Models - such as OpenAI's ChatGPT, and Mistral's Le Chat), including third-party options through community plugins.
Multi-Modal Support: Interacting with an ADK Agent is not simply via a text interface. If required, speak to it and get an audio response. Upload images for it to analyse.
Built-In Evaluation: Critical for enterprise adoption, ADK includes robust testing and evaluation capabilities, ensuring your agents perform reliably before deployment.
Real Business Problems, Real ADK Solutions
So what can you actually DO with this technology? Let's get concrete with some examples:
Customer Service Automation: Create agents that understand customer inquiries across multiple channels, access knowledge bases, update account information, and escalate to human agents when necessary—all while maintaining context throughout the conversation.
Supply Chain Optimisation: Deploy agents that monitor inventory levels, predict shortages, suggest reordering timelines, and adapt to disruptions by suggesting alternative suppliers or logistics routes.
Financial Analysis Assistant: Build agents that gather data from multiple sources, generate reports, identify anomalies, and answer executive questions about financial performance in natural language.
Compliance Monitoring: Implement agents that continuously track regulatory changes, flag potential compliance issues in internal documents, and suggest necessary updates to policies.
Example: logistics and sustainability planner
Following is an example of an AI Agent application that I built using ADK. Its main job is to help figure out the best ways to transport goods while also considering the impact on the environment. Here's what it can do:
Estimate Travel Distances: If you need to send something from City A to City B, the agent can calculate how far it is by road, air, or sea.
Calculate Carbon Footprint: It can estimate the amount of pollution (specifically carbon dioxide, or CO2) that a particular shipment will create. This depends on how it's transported (truck, plane, ship), how far it's going, and how much it weighs.
Compare Options: It can take different shipping options (e.g., sending by air versus by sea) and show you a comparison, often in a chart. This helps you see, for example, which option is faster versus which one creates less pollution. It can even put together options that are, for example, part-road and part-sea.
Find Optimal Routes (Potentially): It has a function that aims to find the "best" shipping route based on what's most important to you – whether that's low cost, speed, or low emissions. (Though, in our recent work, this part was using sample data).
How it does it work?
Think of the AI application as having a toolkit and a set of AI Agents. The Master AI Agent acts as the overarching controller:
The Toolkit:
Specialised Tools: this contains a range of digital tools for different tasks. For instance, one tool is specifically for calculating distances using an online map service. Another tool is like a calculator that uses known figures (called "emission factors") to work out the carbon footprint for different types of transport. A third tool can draw charts.
The Master AI Agent:
Understanding Requests: When it is asked to do something, like "What's the carbon footprint for shipping 1000kg by truck from London to Paris?", the Master AI Agent assesses whether (part or all of) this request is best handled by a another Agent. For example, this example includes a 'Carbon Analyst' Agent which is described as 'an expert in sustainable logistics and carbon accounting.' The Carbon Analyst Agent has a number of specialised tools at its disposal. .
Using the Right Tools: Based on the request, it selects the appropriate tools from its toolkit and feeds them the information they need, like "truck," "London to Paris," "1000kg").
Processing and Responding: The tools do their work (e.g., calculate the distance, then calculate the emissions), and the Master Agent determines if the initial request has been satisfied. If it hasn't (eg because of an error) it attempts to fix the problem, which may require further calls to other agents and tools. Once it believes it has all the required relevant information, it brings it together as requested (eg as a chart) and provides the result.
It's not just one program; it's an AI ecosystem system where a Master Agent orchestrates various specialised functions and data sources to provide helpful insights into carbon footprints.
The Google ADK provides a useful web interface that can be used to test and analyse how the various AI components are working together (or not!) Here are screenshots of showing the flow as a relatively complex question is asked. What you see is the output of Google 'adk web' running on my local computer. The white oval boxes show the names of the tools being called as the Master Agent and other agents make their way through the problem.
The request:
