Grumpy (Google ADK & agents-cli)
An experimental sandbox built to test Google's agents-cli harness and Google ADK capabilities, featuring strict persona enforcement and automated evaluation test sets.
Problem & Context
When exploring new agent harnesses, reading documentation only goes so far. Grumpy was created as a fun, zero-stakes weekend playground specifically designed to stress-test Google's agents-cli and the Google Agent Development Kit (ADK).
The goal wasn't to build a production SaaS or a serious enterprise utility, but to dive under the hood of agent tooling: learning how system instructions are compiled, how multi-turn context is managed in the CLI, how tool bindings function, and how to build a quick LLM-as-a-Judge evaluation loop to measure prompt adherence.
Architecture & How It Works
Grumpy is a lightweight, command-line-driven experiment powered by Gemini Flash via Google's developer toolchain.
[ CLI Input / Test Code ] ──> ( agents-cli & Google ADK ) ──> ( Gemini Flash ) ──> [ Grumpy Output / Eval Suite ]
Playground Harness
Built to explore agents-cli execution workflows, configuration options, and local debugging patterns.
Persona Sandbox
Utilized a comedic, hyper-critical "grumpy reviewer" persona as a stress test to see how aggressively an LLM can be constrained to avoid sycophancy, pleasantries, and code generation.
Eval Harness
A simple Python evaluation script that runs batch prompts against the agent to verify whether prompt instructions hold up across iterative turns.
Key Features
-
CLI Harness Exploration
Hands-on testing of Google ADK structure and local execution via agents-cli.
-
Persona Stress-Testing
Deliberate experimentation with hard negative constraints (e.g., banning apologies, praise, and auto-generated fixes).
-
Automated Test Scripter
Quick evaluation suite checking output consistency against specific behavioral criteria.
Results & Metrics
Successfully mapped the entire lifecycle of building, configuring, and executing custom agents using Google ADK and agents-cli.
Achieved 100% adherence on experimental test prompts within the local sandbox environment.
Tech Stack Deep Dive
- Google ADK
- Framework used for defining agent structure and tool boundaries.
- agents-cli
- Command-line harness used for local agent execution and orchestration.
- Gemini Flash
- Fast, lightweight LLM engine used for rapid prototyping and response generation.
- Python
- Language used for scripting the test harness and evaluation loops.
Links & Resources
Lessons Learned
- 01
CLI Harnesses Accelerate Prototyping
Exploring raw developer tools like agents-cli provides much deeper insight into agent runtime behavior than abstract high-level frameworks.
- 02
Fun Projects Drive Deep Learning
Building a humorous, low-pressure sandbox is often the fastest way to master a new SDK because it removes the friction of worrying about production constraints.