Welcome to Agentic Engineering, Part II
This is a part 2 of a multi-part series on Agentic Engineering. Our family office has been building with AI and investing in it for the last 3 years, this series shares our experience.
Theory without execution is a hallucination.
In Part I, you played the game of simulating an agent. You manually acted as the actuator. Now you must build the actual machine.
Most developers drown in bloated frameworks. They string together fragile abstractions and call it artificial intelligence. They build brittle toys.
You are going to build an engine of leverage. But you are going to start slow!
The industry is infected with a massive misconception. Developers believe they need thousand line libraries to manage a simple API call. They surrender control to the framework. They lose visibility into the execution loop. When the system breaks, they have absolutely no idea why.
You will not make this mistake. You will use Google’s Agent Development Kit. The ADK is not a bloated abstraction layer. It is a precise surgical instrument. It forces you to understand the exact mechanics of the ReAct loop while removing the boilerplate that slows you down. It provides the required primitives.
It enforces the correct architecture.
It gets out of your way.
The ADK Trinity
Complexity is the ultimate enemy of execution. You must organize your system around immutable primitives. The ADK relies entirely on a foundational trinity. You must master the Agent, the Session, and the Runner.
These three components dictate the entire lifecycle of autonomous execution.
The Agent is your logic constraint. It is the definition of the brain. You do not just pass a model a prompt. You bind a specific base model to a strict set of instructions and a curated list of tools. The Agent object is the immutable identity of your autonomous worker. It dictates what the model is allowed to think about. It defines the exact boundaries of its operational authority.
The Session is your ledger of truth. Language models are entirely stateless. They possess the memory of a goldfish. The Session solves this by persisting the exact chronological sequence of every thought, every action, and every observation. It manages the context window ruthlessly. It ensures the Agent never loses its place in the complex execution loop.
The Runner is the heartbeat of execution. The Agent knows the rules and the Session remembers the past, but the Runner actually turns the crank. It takes the user input, queries the Agent, and manages the continuous loop of tool calling. When the Agent outputs a JSON payload requesting data, the Runner intercepts it. The Runner executes your local code. The Runner feeds the result back into the Session.
Why do standard frameworks fail to scale?
They blur the lines between these three vital components.
They mix state management with execution logic. By enforcing the ADK Trinity, you guarantee absolute separation of concerns. You build systems that can scale infinitely without collapsing under their own weight.
CLI & Environment Setup
Your environment dictates your output.
A chaotic workspace breeds chaotic code.
You must establish control immediately.
The command line is your arena for establishing this control.
You initialize your leverage with a single command. You type adk create and you hit enter. This is not a suggestion. This is the only way to establish the proper scaffolding. The ADK CLI generates a pristine workspace. It builds the directory structure. It configures the dependency files. It sets up the exact environment required to run an autonomous loop reliably.
Amateurs build agents in single unmaintainable scripts. They mix their API keys with their business logic. The adk create command forces you into an enterprise grade pattern from day one. You will find a dedicated folder for your tools. You will find a dedicated configuration file for your models. You will find a clean entry point for your core application.
You must respect the structure. The architecture exists to protect you from your own worst impulses. You populate your environment variables with strict discipline. You define your model parameters with absolute intent. You prepare the arena before you ever invite the gladiator into the ring.
Take a hard look at the generated adk.yaml file. This is your mission control. It dictates your runtime environment. It specifies the base model. It locks in your execution parameters. You do not hardcode these variables in your application logic. You manage them centrally.
The Local Dev UI
Blind execution of probabilistic agents is a death sentence.
You cannot debug an autonomous agent by staring at scrolling print statements in a terminal window. You need a microscope. You must observe the exact cognitive steps the model takes before it commits a physical action.
This is where you deploy the adk web command. Executing this command launches a localized browser interface. It completely exposes the inner workings of your system. You stop guessing what the model is doing. You start observing objective reality.
The interface visualizes the entire ReAct paradigm in real time. You submit a prompt. You watch the model generate a thought process. You inspect the exact JSON payload it constructs to call a tool. You see the raw data your system feeds back as an observation. Every single token is accounted for.
How do you fix a hallucinating agent? You look at the UI. You pinpoint the exact moment the logic diverged. You realize your tool description was vague. You realize your prompt lacked specific constraints. You fix the root cause and you run the execution again.
The adk web interface is the ultimate feedback loop for high agency iteration. It tracks your latency. It counts your tokens. It highlights schema validation failures in bright red. It forces you to confront the reality of your code.
You do not deploy to production until the UI proves your loop is bulletproof.
Built-in Tools
An isolated model is entirely useless. It can only recite the historical data it was trained on. To create tangible value you must connect the reasoning engine to current reality. You must arm your agent with actuators.
The ADK provides a robust library of built-in tools. You do not need to invent the wheel for basic internet operations. You need to leverage existing solutions to accelerate your deployment cycle. The GoogleSearchTool is your first weapon of choice. It grants your agent immediate access to the open internet.
Attaching this tool takes a single line of code. You import the module. You append it to the Agent definition array. You instantly transform a static text predictor into a dynamic research assistant. The model now inherently knows it can fetch live information.
This fundamentally alters the prompt dynamic. You no longer ask the model for memorized facts. You assign the model a concrete objective. The model realizes its internal knowledge is insufficient to meet the criteria. It constructs a search query. It formats the JSON arguments. The Runner executes the actual search protocol. The model reads the live results and synthesizes the final report.
You have effectively automated the entire research phase of your workflow.
You did not write web scraping scripts.
You did not manage search API pagination.
You attached a standard actuator and let the reasoning engine handle the rest.
This is the exact definition of engineering leverage.
Practical Exercise: Build a Grounded Assistant
Theory is over. You will now write the code. You will build a Grounded Assistant. You will experience the power of the ADK Trinity firsthand.
Open your workspace. You must define your Agent object first. Select your preferred base model and bind the GoogleSearchTool to its capabilities. Write a system prompt that dictates the model must verify every single claim through a web search before generating a final response. Force the cognitive pause.
Next, you must initialize your Session object. This requires no complex database configuration. You simply instantiate the memory ledger. You give the Agent a place to record its history.
Finally, you define your Runner. Pass it the Agent and the Session objects. You have now assembled the engine. The brain, the memory, and the execution loop are locked together.
Execute the script via the terminal using the adk run command. Do not ask it a simple math question. Ask it about a complex current event that occurred hours ago. Ask it for a specific financial data point that does not exist in any static training corpus.
Hit enter and watch the terminal output.
See the hesitation?
That is the model reasoning.
It is realizing it lacks the data to satisfy your query. It is generating the tool call payload. It is fetching the live web results in the background. It is processing the new context. Suddenly the terminal streams the exact correct answer based on data published today.
You have just built a system of leverage. You took a static language model and forced it to act upon the world. You utilized intensity to learn the framework. You utilized consistency to write the code.
You utilized resilience to debug the loop.
You are now a builder of autonomous systems. You command the tools. You engineer the prompts. You scale the execution. Multiply this leverage infinitely.
Next, in Part III we go deeper down the rabbit hole of workflow control and other agentic engineering mechanics.
Friends: in addition to the 17% discount for becoming annual paid members, we are excited to announce an additional 10% discount when paying with Bitcoin. Reach out to me, these discounts stack on top of each other!
Thank you for helping us accelerate Life in the Singularity by sharing.
I started Life in the Singularity in May 2023 to track all the accelerating changes in AI/ML, robotics, quantum computing and the rest of the technologies accelerating humanity forward into the future. I’m an investor in over a dozen technology companies and I needed a canvas to unfold and examine all the acceleration and breakthroughs across science and technology.
Our brilliant audience includes engineers and executives, incredible technologists, tons of investors, Fortune-500 board members and thousands of people who want to use technology to maximize the utility in their lives.
To help us continue our growth, would you please engage with this post and share us far and wide?! 🙏