Agentic Starter Kits

Purpose

Production-ready starter kits for building and deploying AI agents on Red Hat OpenShift. Each kit is self-contained with docs and deployment configs - most include a Makefile, Dockerfile, and Helm chart. Build locally, then deploy to OpenShift without stitching together boilerplate.

Agents

Agents are organized by framework. Pick one and follow its README.

Templates

Reusable starter kits for building new agents.

Framework	Agent	Description
LangGraph	ReAct Agent	General-purpose agent using a ReAct loop: it reasons and calls tools (e.g. search, math) step by step. Built with LangGraph and LangChain.
LangGraph	Agentic RAG	RAG agent that indexes documents in a vector store (Milvus) and retrieves relevant chunks to augment the LLM's answers with your own data.
LangGraph	ReAct + DB Memory	ReAct agent with PostgreSQL-backed conversation memory. Full chat history is persisted in the database while a FIFO sliding window keeps only the last N messages in the LLM context.
LangGraph	Human-in-the-Loop	ReAct agent with a human approval step. The agent pauses before executing tool calls and waits for user confirmation, enabling oversight of critical actions.
LlamaIndex	WebSearch Agent	Agent built on LlamaIndex that uses a web search tool to query the internet and use the results in its answers.
CrewAI	WebSearch Agent	CrewAI-based agent with a web search tool to query the internet and answer user questions.
Vanilla Python	OpenAI Responses Agent	Minimal agent with no framework: only the OpenAI Python client and an Action/Observation loop with tools. Use with OpenAI or any compatible API.
AutoGen	MCP Agent	AutoGen AssistantAgent with MCP tools over SSE (e.g. churn prediction, math tools), FastAPI /chat/completions.
Google ADK	ADK Agent	General-purpose agent using Google ADK 2.0 with LiteLLM to route inference through an OGX-compatible endpoint.
Langflow	Simple Tool Calling Agent	Tool-calling agent built with Langflow's visual flow builder. Calls external APIs as tools and reasons over results. Includes Langfuse v3 tracing. Runs locally via podman-compose.
A2A	LangGraph + CrewAI Agent	Multi-agent system using the Agent-to-Agent (A2A) protocol. A LangGraph orchestrator delegates tasks to a CrewAI worker agent. Uses a dedicated Helm chart.

Examples

Business use-case demos built on the templates above.

Framework	Agent	Description
LangGraph	CI Failure Summarizer	Spike agent that ingests QG4 GitHub Actions failures, groups incidents in PostgreSQL, composes triage summaries, and posts to Slack via manual POST /summarize trigger.
LangGraph	Guardrailed Agent	Banking agent with NeMo Guardrails safety layer. Content safety, topic boundaries, and regex filtering via the proxy pattern. Two profiles: local (self-check) and nemoguard (supports dedicated NIM classifiers per layer).

Deployments

Pre-built agents and tools deployed on OpenShift.

Framework	Agent	Description
Claude Code	Claude Code on OpenShift	Deploy Claude Code on OpenShift with multiple backend options (Anthropic API, Vertex AI, vLLM, OGX). Includes deployment manifests and configuration guides.
OpenClaw	OpenClaw on OpenShift	Deploy OpenClaw on OpenShift with vLLM model serving and token-based gateway auth (OAuth SSO available via the optional claw-installer path). Kustomize-based deployment using pre-built images.
Codex	Codex on OpenShift	Run OpenAI Codex CLI inside an OpenShell sandbox on OpenShift. Containerfile-based deployment.
OpenCode	OpenCode on OpenShift	Run OpenCode inside an OpenShell sandbox on OpenShift. Containerfile-based deployment.

Getting Started

git clone https://github.com/red-hat-data-services/agentic-starter-kits
cd agentic-starter-kits

Install uv (Python package manager):

curl -LsSf https://astral.sh/uv/install.sh | sh

Pick an agent from the table above, cd into its directory, and follow its README for setup, local development, and deployment to OpenShift.

Browse kits on the starter kit homepage (cards link back to each README).

• Local development — run agents on your machine with Ollama or OGX for model serving. See the Local Development guide.
• Production deployment — containerized deployment on Red Hat OpenShift with Helm charts, monitoring, and production-grade scaling. See the OpenShift Deployment guide.

Repository Structure

agentic-starter-kits/
├── agents/
│   ├── langgraph/                       # Each framework follows this layout
│   │   ├── templates/                   # Agent templates
│   │   │   ├── react_agent/
│   │   │   ├── agentic_rag/
│   │   │   ├── react_with_database_memory/
│   │   │   └── human_in_the_loop/
│   │   ├── examples/                    # Business use-case demos
│   │   │   ├── ci_failure_summarizer/
│   │   │   └── guardrailed_agent/
│   │   └── deployment/                  # Helm chart for this framework
│   ├── crewai/
│   ├── llamaindex/
│   ├── vanilla_python/
│   ├── autogen/
│   ├── google/
│   ├── langflow/
│   ├── claude-code/
│   ├── openclaw/
│   ├── codex/
│   ├── opencode/
│   └── a2a/
├── components/                          # Shared reusable packages (auth, etc.)
├── evals/
│   └── harness/                         # Shared eval engine (runner, scorers, MLflow client)
├── tests/
│   └── behavioral/                      # Behavioral eval suite
├── sandboxes/
│   └── base/                            # OpenShell base container image
├── infrastructure/
│   └── llm-d/                           # llm-d deployment manifests and test tooling
├── docs/                                # Guides: local dev, deployment, contributing
├── pyproject.toml                       # Test deps & pytest config
└── README.md

Behavioral Tests

Behavioral eval suite that tests agents over HTTP against their shared OpenAI-compatible API. Tests are organized by capability so they apply to any agent.

Tests require a running agent. Set the target URL via environment variables:

Env var	Test scope
AGENT_URL	Cross-agent tests (api_contract, adversarial)
REACT_AGENT_URL	LangGraph ReAct agent tests
VANILLA_PYTHON_AGENT_URL	Vanilla Python agent tests
AUTOGEN_MCP_AGENT_URL	AutoGen MCP agent tests
CREWAI_WEBSEARCH_AGENT_URL	CrewAI Websearch agent tests
AGENTIC_RAG_AGENT_URL	LangGraph Agentic RAG agent tests
DB_MEMORY_AGENT_URL	LangGraph DB Memory agent tests
CI_FAILURE_SUMMARIZER_AGENT_URL	LangGraph CI Failure Summarizer agent tests
LLAMAINDEX_WEBSEARCH_AGENT_URL	LlamaIndex Websearch agent tests
LANGFLOW_TOOL_CALLING_AGENT_URL	Langflow Simple Tool Calling agent tests
LANGFLOW_FLOW_ID	Langflow flow ID (changes on re-import)
HITL_AGENT_URL	LangGraph Human-in-the-Loop agent tests
GUARDRAILED_AGENT_URL	LangGraph Guardrailed agent tests
GOOGLE_ADK_AGENT_URL	Google ADK agent tests
A2A_LANGGRAPH_CREWAI_AGENT_URL	A2A LangGraph-CrewAI agent tests

uv pip install -e ".[test]"
AGENT_URL=https://my-agent.example.com pytest tests/behavioral/ -v

See tests/behavioral/ for full details.

Documentation

• Local Development — Ollama + OGX setup
• OpenShift Deployment — Helm-based deployment guide
• Adding a New Agent — How to contribute a new agent template
• Adding Behavioral Tests — How to add test coverage for an agent
• Guardrails — Adding safety rails to agents (NeMo Guardrails vs Guardrails Orchestrator)
• llm-d Deployment — Deploy llm-d for intelligent LLM inference routing on OpenShift AI

Claude Code Skills

This project has Claude Code skills that automate contributor workflows — tracing integration, deployment, test scaffolding, and more. Contributor and operator skills are in the agentic-starter-kits-skills plugin. See CONTRIBUTING.md — Claude Code skills for installation instructions and the full skill list.

Additional Resources

• OGX Documentation: https://ogx-ai.github.io/docs
• Ollama Documentation: https://docs.ollama.com/
• OpenShift Documentation: https://docs.openshift.com/
• Kubernetes: https://kubernetes.io/docs/home/

Contributing

Contributions are welcome! See CONTRIBUTING.md for guidelines.

License

Apache License 2.0