🤖 DevOps Pipeline Failure Agent

An agentic AI that automatically reads a failed CI build log, inspects the source code, diagnoses the root cause, and posts a fix recommendation to Slack — all autonomously.

Two versions: Claude API (cloud) and Qwen3 via Ollama (100% local, zero cost).

📺 Video Series

Video	Title	Link
Part 1	Agentic AI in DevOps — The Skill Every Cloud Engineer Needs in 2026	Watch
Part 2	Local LLM vs Cloud API for DevOps Agents — Which Should You Use in 2026?	Watch
Part 3	Wire Your AI Agent to Real GitHub Actions	🔜

Cloud Advocate · GK

🧠 How It Works

Traditional DevOps automation executes what you tell it to execute. Agentic AI reasons, decides, and acts on its own.

Pipeline fails
     ↓
Agent reads the build log              ← Perceive
     ↓
LLM identifies the failing             ← Plan
test and decides which file to inspect
     ↓
Agent reads the source code            ← Act
     ↓
LLM diagnoses root cause              ← Observe
     ↓
Fix recommendation posted to Slack     ← Report
     ↓
Loop repeats until job is done         ← ↻ Agentic loop

The key insight: The LLM is the manager. The tools are the workers. The manager never does the work itself — it delegates to the right worker at the right time.

⚡ Two Agents — Choose Your Setup

	`agent.py`	`agent_local.py`
Model	Claude Sonnet (Anthropic)	Qwen3:14b (Ollama)
Cost	~$0.01 per run	$0.00 forever
Privacy	Cloud API	100% local
Internet	Required	Not required
Setup	API key only	Ollama + GPU
Diagnosis	✅ Correct	✅ Correct
Enterprise safe	Check policy	Always safe

🔵 Option 1 — Claude API (agent.py)

Prerequisites

Python 3.12+
WSL2 / Linux / macOS
Anthropic API key → console.anthropic.com

Setup

# 1. Clone the repo
git clone https://github.com/cloudadvocate/devops-agent.git
cd devops-agent

# 2. Create virtual environment
python3 -m venv .venv
source .venv/bin/activate

# 3. Install dependencies
pip install -r requirements.txt

# 4. Set your API key
export ANTHROPIC_API_KEY=sk-ant-your-key-here

Run

python agent.py

# Or use your own CI log
python agent.py --log /path/to/build.log

🟢 Option 2 — Local Qwen3 via Ollama (agent_local.py)

Zero API cost. Zero data leaves your machine.

Prerequisites

Python 3.12+
WSL2 / Linux / macOS
Ollama installed on your machine
GPU recommended (see hardware table below)

Hardware recommendations

VRAM	Recommended model	Pull command
CPU only	qwen3:8b (slow)	`ollama pull qwen3:8b`
8GB	qwen3:8b	`ollama pull qwen3:8b`
12GB	qwen3:14b ← sweet spot	`ollama pull qwen3:14b`
16GB	qwen3:30b-a3b	`ollama pull qwen3:30b-a3b`
24GB	qwen3:32b	`ollama pull qwen3:32b`

Setup

# 1. Install Ollama → https://ollama.com
# 2. Pull the model
ollama pull qwen3:14b

# 3. Start Ollama server (keep this running)
ollama serve

# 4. Clone the repo and set up venv
git clone https://github.com/cloudadvocate/devops-agent.git
cd devops-agent
python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

WSL2 users — set Ollama host

Ollama runs on Windows. WSL2 needs to point to it:

# Find your Windows host IP
ip route | grep default | awk '{print $3}'

# Set the host (replace with your IP)
export OLLAMA_HOST=http://172.24.48.1:11434

# Verify connection
curl http://172.24.48.1:11434/api/tags

Run

export OLLAMA_HOST=http://YOUR_WINDOWS_IP:11434
python agent_local.py

🎬 Demo Output

Claude API version

════════════════════════════════════════════════════════════
   🤖  DevOps Pipeline Failure Agent
════════════════════════════════════════════════════════════
   Model: claude-sonnet-4-20250514
   Log:   samples/failed_build.log

⟳  Agent thinking... (step 1)
🔧  Calling tool: read_build_log
⟳  Agent thinking... (step 2)
🔧  Calling tool: read_source_file  (file_path='gateway.py')
⟳  Agent thinking... (step 3)
🔧  Calling tool: post_slack_summary

────────────────────────────────────────────────────────────
  📬  Slack #devops-alerts
────────────────────────────────────────────────────────────
  🔴  Build Failure Diagnosed
  Root cause:  Timeout threshold (5s) too low for Stripe API
  Affected file:  gateway.py · Severity: HIGH
  Fix:  Increase DEFAULT_TIMEOUT to 15s
  _Diagnosed by Cloud Advocate DevOps Agent · 14:23:01_
────────────────────────────────────────────────────────────
  Tokens used — input: 2,847  output: 312
════════════════════════════════════════════════════════════

Local Qwen3 version

════════════════════════════════════════════════════════════
   🤖  DevOps Agent — Local Edition (Qwen3:14b)
════════════════════════════════════════════════════════════
   Model:  qwen3:14b via Ollama (100% local)
   Cost:   $0.00

⟳  Agent thinking... (step 1)
🔧  Calling tool: read_build_log
⟳  Agent thinking... (step 2)
🔧  Calling tool: read_source_file  (file_path='gateway.py')
⟳  Agent thinking... (step 3)
🔧  Calling tool: post_slack_summary

────────────────────────────────────────────────────────────
  📬  Slack #devops-alerts
────────────────────────────────────────────────────────────
  🔴  Build Failure Diagnosed
  Root cause:  PaymentGateway returns timeout instead of raising
  Affected file:  gateway.py · Severity: HIGH
  Fix:  raise PaymentGatewayTimeout() instead of returning result
  _Diagnosed by Cloud Advocate Local Agent (Qwen3:14b) · 08:40:34_
────────────────────────────────────────────────────────────
  Powered by: qwen3:14b running locally via Ollama
════════════════════════════════════════════════════════════

🏃 Part 3 — Real GitHub Actions on a Self-Hosted Runner

Parts 1 & 2 ran the agent by hand against a canned log. Part 3 wires it into a real GitHub Actions pipeline: a real pytest suite fails on your own Ubuntu machine (registered as a self-hosted runner), and that failure automatically triggers agent_local.py — fully local, zero API cost, diagnosis posted straight to the Actions job summary.

git push
     ↓
GitHub Actions dispatches the job to YOUR Ubuntu runner
     ↓
pytest tests/ runs for real → test_payment_retry_logic fails
     ↓
if: failure() triggers agent_local.py --log ci_output.log
     ↓
Qwen3:14b (via Ollama, same box) reads the log + gateway.py
     ↓
Diagnosis + fix written to $GITHUB_STEP_SUMMARY and printed to the job logs

1. Prerequisites on the runner box

Ollama installed and running (ollama serve), with qwen3:14b pulled — see Option 2 above.
Python 3.12+ and git.
This repo pushed to your own GitHub remote (not a fork you don't control).

2. Register the self-hosted runner

In your GitHub repo: Settings → Actions → Runners → New self-hosted runner → pick Linux/x64. GitHub shows you a one-time download + config script for your repo, e.g.:

mkdir actions-runner && cd actions-runner
curl -o actions-runner-linux-x64.tar.gz -L https://github.com/actions/runner/releases/download/vX.Y.Z/actions-runner-linux-x64-X.Y.Z.tar.gz
tar xzf ./actions-runner-linux-x64.tar.gz
./config.sh --url https://github.com/YOUR_USER/devops-agent --token YOUR_REGISTRATION_TOKEN

Run it as a background service so it survives reboots and keeps listening for jobs:

sudo ./svc.sh install
sudo ./svc.sh start

(Or ./run.sh in a terminal you keep open, if you'd rather watch it live for the demo.)

3. Trigger it

Push to main, open a PR, or trigger it manually from the Actions tab (workflow_dispatch). Watch the run: pytest fails on test_payment_retry_logic, and the Diagnose failure with local AI agent step kicks in — its output and the job summary show the same root cause the Part 1/2 demos surfaced, but this time diagnosed from a log that was generated by a real, live test run on your own hardware.

4. The workflow

See .github/workflows/ci.yml — no secrets required for this path, since everything (tests, model, diagnosis) runs on your own box.

📁 Project Structure

devops-agent/
├── agent.py                   # Cloud agent — Claude API
├── agent_local.py             # Local agent — Qwen3 via Ollama
├── requirements.txt           # Dependencies
├── samples/
│   ├── failed_build.log       # Canned sample log (used for the standalone demo)
│   └── gateway.py             # Buggy source file the agent inspects
├── tests/                     # Real pytest suite — one test genuinely fails
│   ├── conftest.py
│   ├── test_payments.py       # The failing test (mocked Stripe timeout)
│   ├── test_api.py
│   ├── test_auth.py
│   ├── test_health.py
│   └── test_utils.py
└── .github/
    └── workflows/
        └── ci.yml             # Real GitHub Actions trigger on a self-hosted runner

🛠️ The Three Tools

The agent has exactly three tools — think of them as workers it can delegate to:

Tool	What it does
`read_build_log`	Reads the CI failure output — stack traces, error messages, test results
`read_source_file`	Reads any source file from the repo to inspect the buggy code
`post_slack_summary`	Posts the diagnosis and fix recommendation to Slack

The LLM (the manager) decides which tool to call, when, and in what order. You never hardcode that logic — the model reasons its way to it.

🔍 The Agentic Loop

Claude API version (agent.py)

while True:
    response = client.messages.create(
        model=MODEL, tools=TOOLS, messages=messages
    )
    if response.stop_reason == "end_turn":
        break
    for block in response.content:
        if block.type == "tool_use":
            result = execute_tool(block.name, block.input)
            messages.append({"role": "assistant", "content": response.content})
            messages.append({"role": "user", "content": [tool_result]})

Local Ollama version (agent_local.py)

while step < MAX_STEPS:
    response = ollama.chat(
        model=MODEL, messages=messages,
        tools=TOOLS, think=False, options={"temperature": 0.1}
    )
    if not response.message.tool_calls:
        break
    for tool_call in response.message.tool_calls:
        result = execute_tool(tool_call.function.name, tool_call.function.arguments)
        messages.append({"role": "tool", "content": result})

Key difference: Ollama uses OpenAI-compatible tool format. Claude uses Anthropic format. The agentic loop logic is identical — only the API client changes.

🚀 Extending This Agent

Add a real Slack webhook

import requests

requests.post(
    os.environ["SLACK_WEBHOOK_URL"],
    json={"text": f"*{payload['root_cause']}*\nFile: `{payload['affected_file']}`\n```{payload['code_suggestion']}```"}
)

Add more tools

{"name": "create_github_pr",      "description": "Create a PR with the fix"}
{"name": "query_datadog_metrics",  "description": "Fetch metrics around failure time"}
{"name": "run_terraform_plan",     "description": "Run terraform plan and return diff"}

Wire to GitHub Actions

.github/workflows/ci.yml does this for real on a self-hosted runner with the local Qwen3 agent — see Part 3 above.

Prefer the cloud agent instead (e.g. on a GitHub-hosted runner, no local Ollama box)? Swap the last step for:

- name: Diagnose failure with Claude
  if: failure()
  env:
    ANTHROPIC_API_KEY: ${{ secrets.ANTHROPIC_API_KEY }}
  run: python agent.py --log ci_output.log

🔒 Production Guardrails

Guardrail	Why
Dry-run mode	Agent shows what it would do before acting
Approval gate	Human approves fix in Slack before PR is raised
Scope limits	Agent cannot touch production without explicit flag
Audit logging	Every tool call logged with timestamp and reasoning
Secret scanning	Never pass API keys through agent context

📦 Requirements

anthropic>=0.40.0
ollama>=0.4.0
requests>=2.31.0

Version	Cost per run
Claude API	~$0.01–0.02
Qwen3 local	$0.00

🗺️ Roadmap

Claude API agent (Part 1)
Local Qwen3 agent via Ollama (Part 2)
GitHub Actions automatic trigger on build failure (Part 3)
Self-hosted runner setup guide (Part 3)
Real Slack webhook integration
GitHub PR creation tool
Multi-agent pattern — planner + executor + reviewer
Datadog / CloudWatch metrics correlation

👨‍💻 About

Built by GK — Cloud Architect and DevSecOps practitioner with 10+ years of enterprise networking and telecom experience.

Cloud Advocate is a YouTube channel covering AI on Cloud, DevSecOps, Multi-cloud, FinOps, and Cloud Career guidance with real-world hands-on depth.

📺 YouTube — Cloud Advocate 💼 LinkedIn

📄 License

MIT License — free to use, modify, and distribute.

If this helped you, give the repo a ⭐ and subscribe to Cloud Advocate for more hands-on cloud and DevOps content.

Name		Name	Last commit message	Last commit date
.github/workflows		.github/workflows
samples		samples
tests		tests
.gitignore		.gitignore
README.md		README.md
agent.py		agent.py
agent_local.py		agent_local.py
requirements.txt		requirements.txt

Latest commit

History

Folders and files

Repository files navigation

🤖 DevOps Pipeline Failure Agent

📺 Video Series

🧠 How It Works

⚡ Two Agents — Choose Your Setup

🔵 Option 1 — Claude API (agent.py)

Prerequisites

Setup

Run

🟢 Option 2 — Local Qwen3 via Ollama (agent_local.py)

Prerequisites

Hardware recommendations

Setup

WSL2 users — set Ollama host

Run

🎬 Demo Output

Claude API version

Local Qwen3 version

🏃 Part 3 — Real GitHub Actions on a Self-Hosted Runner

1. Prerequisites on the runner box

2. Register the self-hosted runner

3. Trigger it

4. The workflow

📁 Project Structure

🛠️ The Three Tools

🔍 The Agentic Loop

Claude API version (agent.py)

Local Ollama version (agent_local.py)

🚀 Extending This Agent

Add a real Slack webhook

Add more tools

Wire to GitHub Actions

🔒 Production Guardrails

📦 Requirements

🗺️ Roadmap

👨‍💻 About

📄 License

About

Resources

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages