Google's Agents CLI: Build & Deploy Agents in Minutes
Shubham Saboo demos Agents CLI for scaffolding, evaluating, and deploying AI agents via simple terminal prompts, handling configs and cloud setup automatically.
User Understanding Trumps Model Complexity in Agent Success
Shubham Saboo, creator of the 105K-star Awesome LLM Apps GitHub repo, traces AI agents' evolution from GPT-3's prompt engineering era—where afternoons were spent crafting JSON outputs—to today's structured outputs via Pydantic schemas that "just work." Back then, agents were "janky loops around a completion call with string parsing"; now, he runs six agents on cron jobs handling daily tasks.
What hasn't changed: success hinges on user and problem comprehension. "The model is a universal function now... table stakes. Everybody has the model," Saboo says. Winners shape problems clearly for the model, communicate effectively with agents, and treat them like interns for optimal output. Host Smitha Kolan notes persistent skills like user focus amid tech shifts.
Saboo's repo started as personal organization for local GPT-3 experiments but exploded after hitting 1,000 stars in weeks, revealing demand for runnable samples. It now ranks in GitHub's top 100 repos, landing him a Google PM role. Lesson: Build publicly to solve your pains; others follow.
"Your prompt is as good as your understanding of the problem... that's even more true now because everybody has access to these models and agents."
Agents CLI Handles Full Agent Lifecycle from English Prompts
Google's Agents CLI, paired with skills packages, equips coding agents (Gemini CLI, Claude, Cursor) to build, eval, and deploy ADK agents without hallucinations or manual YAML/config hell. Install via one uvx command; it auto-scaffolds projects, sets environments, and integrates ADK knowledge.
Demo 1: "Caveman Compressor"—verbose text to grunts. Prompt Gemini CLI: "Use agent CLI to build a caveman style agent that compresses verbose text into technical grunts." In <1 minute: scaffolds folder, installs deps, generates code, runs locally via ADK web UI (localhost:8080 chatbot with event logs/states/artifacts). Deploy to Agent Engine (5-10 mins) with explicit approvals, yielding cloud dashboard, traces, playground.
No console switching or doc-pasting needed—CLI manages it all. Kolan highlights skipping ADK docs context in coding agents.
Extend via prompts: Add Google Search tool (internet access), RAG (grounding in docs/DBs), multi-agent workflows. Saboo: "99% of the time in one shot."
"Agent CLI really fixes hallucinations... everything packaged into a single CLI... your coding agents have access to all the internal tools, codebase, and knowledge about ADK."
Evaluations, Multi-Agents, and Production Resilience
Post-build, prompt for evals: "Generate 20 eval criteria for caveman agent and run them." Auto-generates/tests, flags fails for fixes. All passed in demo.
Demo 2: Multi-agent "PR Roaster"—roasts GitHub PRs. Builds graph-based workflow (ADK 2.0 upgrade over prompts), deploys similarly. Live roast pokes fun at Kolan's code.
New ADK features: Graph workflows for complex orchestration; resumable agents survive drop connections (production reality); ambient agents run 24/7 via Agent Engine cron-like scheduling. Multi-language: Python, TS, Go, Java.
Tools integrate seamlessly: Google Search, Cloud Storage, MCPs. Observability (traces, logs) baked in.
Saboo stresses embeddings knowledge: Every developer needs it for RAG/agents. RAG isn't dead—evolves. Soft skills (clear thinking, communication) now core tech requirements.
"I have six agents running on a cron job that does all the work for me... the only limitation now is how creative you can get with it, how clearly you can think about the problem."
Key Takeaways
- Install Agents CLI (
uvxcommand) to supercharge coding agents for ADK: scaffolds, evals, deploys from English prompts—no YAML/config hassle. - Test locally with ADK web UI (chatbot + event logs) before cloud deploy to Agent Engine for production traces/playground.
- Generate/run evals automatically: Prompt coding agent for criteria; flags fails for iteration.
- Extend via prompts: Add tools (Google Search), RAG, multi-agents—handles 99% cases one-shot.
- Build resilient agents: Use resumable flags for dropouts, ambient for 24/7 runs.
- Focus on users/problems over prompts: Model access is table stakes; shape inputs clearly.
- Learn embeddings: Powers every RAG/agent; essential for devs.
- Start simple: Publicly share experiments (like Awesome LLM Apps) to validate demand.
- Multi-lang support (Python/TS/Go/Java) for diverse stacks.
- Treat agents like interns: Clear communication yields best results.