08-ai-agents-for-devops

AI Agents for DevOps (LangGraph + LangChain + Ollama)

Build a local, private AI agent that reads log files and explains what happened, running entirely on your machine with Ollama, orchestrated with LangGraph and LangChain.

The idea that keeps it reliable: the tool does the math, the LLM writes the story. A plain Python function counts the log levels (and is unit-tested without any LLM); the agent just calls that tool and summarizes the result, so the numbers are always right even if a small local model is imperfect at tool calls.

What you'll learn

• What an agent is: an LLM that can decide to call tools
• Wrap a Python function as a LangChain @tool
• Run a local LLM with ChatOllama (no API keys, no cloud)
• Build an agent with create_agent, and see the same loop by hand with a LangGraph StateGraph

Prerequisites

Ollama running with a model pulled:

ollama serve
ollama pull llama3.2                 # default (small, ~2GB)
curl http://localhost:11434/api/tags # should list llama3.2

llama3.2 reliably calls the tool but its summaries can be generic. For sharper output, pull a stronger model and set OLLAMA_MODEL:

ollama pull qwen3-coder:30b
OLLAMA_MODEL=qwen3-coder:30b python run_agent.py

Setup

python3 -m venv venv
source venv/bin/activate
pip install -r requirements.txt

Files

• app.log — sample log
• log_utils.py — deterministic counting core (no LLM)
• run_agent.py — the agent (create_agent, plus a hand-built StateGraph)
• test_log_utils.py — tests for the counting core

python -m pytest .            # counting core, no Ollama needed
python run_agent.py           # the agent (needs Ollama)
python run_agent.py custom    # the hand-built StateGraph version

The demo prints the deterministic counts first, then the agent's summary, so you always see correct numbers. create_agent (from langchain.agents) is the LangChain 1.0 API; run_agent.py custom shows the same agent/tools loop built by hand with StateGraph, ToolNode and a conditional edge.

Practice

Wire up the agent yourself:

1. Create a ChatOllama model (base_url="http://localhost:11434", temperature=0).
2. Pass the analyze_log_file tool to create_agent(model, tools=[...], system_prompt=...).
3. Invoke it with a HumanMessage asking it to analyze app.log, and print the result.
4. Bonus: add a second tool (e.g. return the last N lines) and watch how it chooses.