, fallback_conn_ids=["anthropic_prod"] , ) # rerun the task # ask the model again # try another vendor Fallback: the next connection, inside the same attempt 30
it @task( retries=3, retry_policy=LLMRetryPolicy(llm_conn_id="pydanticai_default") , ) def build_report(): ... Works on any task · Airflow 3.3+ The model reads the error: retry or not, and how long to wait 31
write_draft() judge(draft) # a model writes # a model scores it refinement = refine.loop( max_iterations=10 , until=good_enough) New in Airflow 3.4 · max_iterations is required 34
· released 6 November 2026 Task Loops Graceful pause Beyond Python You just saw them: repeat until good enough Drain a Dag, then it pauses itself Dags in TypeScript, Go or Java 39