Agentic DAG Template

This template implements a production-ready agentic DAG with the OpenAI Agents SDK and Prefect. It includes a minimal wrapper layer that exposes core SDK features (guardrails, sessions, tracing, streaming, MCP approvals, handoffs) using Pydantic configuration.

Quickstart (uv)

1. Install dependencies

uv sync

2. Configure environment

cp .env.example .env

3. Start the local MCP stub (Streamable HTTP)

uv run python -m src.mcp_stub_server

4. Start Prefect UI

prefect server start

5. Run the flow

uv run python -m src.flows.work_order_flow

Open the Prefect UI at http://127.0.0.1:4200 to visualize the DAG.

Project demos

• Streaming + sessions demo:

uv run python -m src.projects.interactive_assistant

Wrapper features

• Guardrails: input/output guardrails and tool guardrails in src/wrapper/guardrails.py.
• Sessions: SQLite-backed sessions via src/wrapper/sessions.py.
• Tracing: controlled by TRACING_* env vars and RunConfig in src/config.py.
• Streaming: helper in src/wrapper/runner.py for Runner.run_streamed().
• Approvals: hosted MCP approval hook in src/wrapper/approvals.py.
• Handoffs: helper in src/wrapper/handoffs.py.
• Pydantic config: src/config.py uses BaseModel and AppConfig.from_env().

Notes

• MCP server endpoint is configured with MCP_URL (defaults to http://localhost:8000/mcp).
• The MCP stub exposes search_docs, read_metrics, create_task, update_ticket tools.
• To enable hosted MCP approvals, set USE_HOSTED_MCP=true and MCP_HOSTED_URL.

Docs references

• OpenAI Agents SDK (agents, handoffs, MCP): https://openai.github.io/openai-agents-python/
• OpenAI Agents SDK repository: https://github.com/openai/openai-agents-python
• Prefect flow run pause/resume: https://reference.prefect.io/prefect/flow_runs/