You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Astro Runtime 3.4 alpha (runtime-dev:3.4-1-alpha6, Airflow 3.4 pre-release) with
apache-airflow-providers-common-ai 1.1.0. Package versions are pinned in
requirements.txt.
How to use this repository
Copy .env_example to .env and add your own credentials for your model provider, adjust the model in AIRFLOW_CONN_PYDANTICAI_DEFAULT if needed.
Start Airflow
astro dev start
Layout
dags/
01_llm_tasks/
decorators/ one Dag per @task.llm* decorator
operators/ classic-operator forms
02_agent/
basics/ @task.agent, from minimal to many parameters combined
durable/ durable=True
capabilities/ capabilities= and CodeMode
03_toolsets/ one Dag per toolset, plus one that combines two
04_hitl/
standard_operators/ human-in-the-loop operators from the standard provider
agent_review/ common-ai's enable_hitl_review
agent_tool_approval/ per-tool approval on @task.agent
05_connections_hooks/ the hook API and the MCP connection transports
06_policies/ LLM-backed retry and decision policies
07_agent_frameworks/ Common AI toolsets in LangChain, OpenAI Agents SDK and Strands agents
08_documents/ document loading and LlamaIndex retrieval
09_task_loops/ an agent inside an Airflow 3.4 task loop
other/ end-to-end pipelines built out of the above
framework_comparison/ the same Dag six times, once per agent framework
01_llm_tasks
Dag
Provider component
example_llm_operator
LLMOperator / @task.llm with a Pydantic output_type and UsageLimits
example_llm_branch
LLMBranchOperator / @task.llm_branch
example_llm_sql_query
LLMSQLQueryOperator / @task.llm_sql
example_llm_file_analysis
LLMFileAnalysisOperator / @task.llm_file_analysis
example_llm_schema_compare
@task.llm_schema_compare
example_llm_batch
@task.llm_batch, the provider's batch API with results written to object storage
HookToolset. Any Airflow Hook's methods can become tools
example_agent_datafusion_toolset
DataFusionToolset
example_agent_mcp_toolset
MCPToolset
example_agent_object_storage_toolset
ObjectStorageToolset
example_agent_skills_toolset
AgentSkillsToolset with a local skill in include/skills
example_agent_sandbox_toolset
SandboxToolset with OpenSandboxBackend, against the local OpenSandbox server
example_agent_logging_toolset
Tool-call logging with enable_tool_logging=True
example_agent_multi_toolset
SQLToolset and HookToolset on one agent
04_hitl
Dag
Operator
Info
example_hitl_approval
ApprovalOperator
Approve or reject. Rejecting skips the downstream tasks
example_hitl_operator
HITLOperator
N options, multiple=True
example_hitl_branch
HITLBranchOperator
options_mapping
example_hitl_entry
HITLEntryOperator
Input form
example_agent_hitl_review
common-ai enable_hitl_review=True
Note: HITL interaction in a separate plugin tab on the task instances
example_agent_tool_approval
common-ai requires_approval=True on a tool
The pending tool call shows up on the Required Actions page
example_agent_hitl_review is the only Dag here that uses common-ai's React review plugin.
The plugin loads its JS bundle from AIRFLOW__API__BASE_URL, so if your Airflow
isn't on http://localhost:8080 you need the base-URL block in
docker-compose.override.yml, which ships commented out. The four standard
operators work either way.
05_connections_hooks
Dag
Provider component
example_pydantic_ai_hook
PydanticAIHook.create_agent()
example_mcp_connection_transports
mcp connection using stdio and http
example_mcp_hook
MCPHook used directly, without an agent
example_langchain_hook
LangChainHook chat and embedding models
example_llm_fallback
fallback_conn_ids: the first connection always fails, so the task uses the backup connection
06_policies
Dag
Provider component
example_llm_retry_policy
LLMRetryPolicy
example_classifier_retry_policy
ClassifierRetryPolicy
example_decision_policy
DecisionPolicy on @task.llm (experimental)
07_agent_frameworks
Dag
Provider component
example_langchain_toolset_bridge
airflow_toolset_to_langchain_tools with a LangChain create_agent agent
example_openai_agents_tools
OpenAI Agents SDK AirflowTools and AirflowTraceProcessor
example_openai_agents_durable
run_durably_sync: a retry replays the completed model and tool calls
example_strands_tools
Strands AirflowTools plugin
example_strands_durable
invoke_durably: a retry resumes from the last checkpoint
Both durable Dags simulate a worker crash on try 1 to show the replay on try 2.
08_documents
Dag
Provider component
example_document_loader
DocumentLoaderOperator on the Markdown files in include/mission_logs
example_llamaindex_rag
LlamaIndexEmbeddingOperator and LlamaIndexRetrievalOperator
09_task_loops
Dag
What it builds
example_agent_text_to_sql_loop
@task.agent in a task loop that retries until its SQL passes a check
other
Dag
What it builds
email_routing
@task.llm sorts mail into P0-P4, @task.branch routes it
support_reply_evals
LLM as judge. Scores agent replies and customer reactions, CSAT
Agent framework comparison
Six Dags run an identical support-ticket flow. Fetch a ticket, draft a
structured TicketResponse using a lookup_shipment tool, HITL review branch.
Dag
Framework
support_ticket_common_ai
common-ai provider (@task.agent)
support_ticket_pydantic_ai
pydantic-ai used directly
support_ticket_langgraph
LangGraph ReAct agent
support_ticket_crewai
CrewAI single-agent crew, in a @task.virtualenv
support_ticket_temporal
Temporal workflow
support_ticket_direct_api
OpenAI SDK and a hand-written tool loop
All examples other than the common AI one use gpt-5-mini with OPENAI_API_KEY, adjust for other model providers.
CrewAI requires openai<3 and the Common AI provider needs openai>=3, so
support_ticket_crewai installs CrewAI and openai 2.x into a virtualenv at
task runtime instead of into the image, which adds install time to every run of
that task.
example_agent_sandbox_toolset uses the opensandbox-server service from
docker-compose.override.yml (OpenSandbox 1.2.0). The server mounts the host's Docker
socket and starts each sandbox as a container on your local Docker, with an egress
sidecar in dns+nft mode so the default deny-all network policy is enforced. The
server publishes no port and has no API key; only containers on the airflow
network reach it. Sandbox ports are published on 127.0.0.1 only.
The Temporal Dag uses the temporal service from docker-compose.override.yml. Its web UI is at http://localhost:8233.