DEV Community

Cover image for Build a CI/CD Agent That Fixes Its Own Failures — in 5 Minutes Without Writing Any Boilerplate
Sushil Kulkarni
Sushil Kulkarni

Posted on

Build a CI/CD Agent That Fixes Its Own Failures — in 5 Minutes Without Writing Any Boilerplate

#devchallenge #cloudnextchallenge #googlecloud #cli

Google Cloud NEXT '26 Challenge Submission

This is a submission for the Google Cloud NEXT Writing Challenge

Everyone talked about "Agentic Enterprise" at Google Cloud NEXT '26.

But the most important release wasn’t the platform, the sandbox, or even Gemini.

It was a CLI.

And once I understood what agents-cli actually does, it completely changed how I think about developer tools.

This article is about that shift — and how I built a self-healing CI/CD agent in minutes… without writing the code myself.

TL;DR

  • Google is building CLIs for AI agents, not humans
  • agents-cli exposes machine-readable skills to coding assistants
  • This removes boilerplate and hallucination issues
  • You describe intent → AI executes real commands
  • This is a fundamental shift in developer tooling

What Happened

At Google Cloud NEXT '26, we saw major announcements:

  • Gemini Enterprise Agent Platform
  • GKE Agent Sandbox
  • Agent Development Kit (ADK) updates

All impressive.

But buried inside those announcements was something much more disruptive:

agents-cli

At first glance, it looks like just another CLI tool.

It’s not.

The Real Shift: CLIs Are Now Built for AI

This is the part most people missed.

Traditionally, CLIs were designed for humans:

  • Memorize commands
  • Read documentation
  • Write boilerplate
  • Debug syntax

But agents-cli flips this model completely.

It turns CLI capabilities into machine-readable skills that AI assistants can:

  • Discover
  • Execute
  • Chain together

Which means:

👉 You don’t use the CLI directly

👉 Your AI uses the CLI for you

This is the real paradigm shift.

The Problem: AI + New Frameworks = Friction

For the last few years, developers have been stuck in a loop:

  1. A new AI framework drops
  2. Docs are incomplete or evolving
  3. You ask your AI assistant for help
  4. It hallucinates or gets syntax wrong

The root issue isn’t just hallucination.

It’s time mismatch.

Frameworks evolve faster than AI models can learn them.

So AI is always slightly outdated.

How agents-cli Fixes This

Instead of relying on training data, agents-cli gives AI assistants live, executable knowledge.

It injects a skills layer into your environment.

So your AI assistant can:

  • Run real commands
  • Scaffold projects correctly
  • Follow actual platform conventions
  • Avoid guessing

This is a huge shift:

From predicting code → to executing capabilities

How I Built a Self-Healing CI/CD Agent in 5 Minutes

Here’s exactly what I did.

Step 1: Inject the Skills 

uvx google-agents-cli setup
Enter fullscreen mode Exit fullscreen mode

This single command connects your local environment to the agent ecosystem.

Think of it as giving your AI assistant real tools instead of guesses.

Step 2: Describe What You Want

Instead of writing code, I opened my terminal-based AI assistant and said:

"Use the google-agents-cli-scaffold skill to create a project called cicd-healer-agent using the prototype flag. Then create an agent that analyzes failing CI/CD logs and outputs a fix."

No boilerplate. No syntax memorization.

Step 3: Let the AI Execute

Because the AI had access to agents-cli skills, it:

  • Ran the correct CLI commands
  • Scaffolded the project properly
  • Generated valid ADK workflow logic
  • Wired everything together

No hallucination. No broken code.

Step 4: Make It Safe with GKE Agent Sandbox

A CI/CD agent that writes code is powerful — but risky.

Running AI-generated code directly on your machine is dangerous.

This is where GKE Agent Sandbox comes in.

It provides:

  • Isolated execution environments
  • gVisor-based security
  • Safe testing of generated patches

So your agent can:

  1. Generate a fix
  2. Test it safely
  3. Propose a patch

Without risking your system.

Real Output: CI/CD Healer Agent in Action

Here’s my CI/CD agent diagnosing a failure and generating a fix — in real time.

CI/CD Healer Agent Output

A New Mental Model: You’re Not Coding Anymore

This is the biggest mindset shift.

We’re moving from:

  • Writing code
    → to describing intent

  • Using tools
    → to equipping AI with tools

The question is no longer:

“Can AI write this code?”

It’s:

“Does my AI have the right capabilities to execute this?”

Practical Takeaways

If you’re building with AI agents today:

  • Stop focusing only on prompts
  • Start thinking in capabilities and tools
  • Prefer systems that give AI real execution power
  • Avoid workflows where AI has to “guess”

And most importantly:

Give your AI structured access to your stack

That’s where tools like agents-cli shine.

Final Thoughts

The era of writing boilerplate is ending.

agents-cli shows us what comes next:

  • Intent-driven development
  • AI-native tooling
  • Programmable agents with real capabilities

Once you experience this workflow, going back feels… slow.

The real question is:

Are you still coding…

Or are you orchestrating?

What Do You Think?

Are tools like agents-cli the future of development?

Or are we giving AI too much control over our workflows?

Curious to hear your thoughts 👇

Top comments (0)