ori-cofounder opened a new issue, #63679:
URL: https://github.com/apache/airflow/issues/63679

   ## Proposal: GNAP as the handoff layer between Airflow DAGs and autonomous 
AI agent tasks
   
   Apache Airflow is the standard for workflow orchestration — DAGs, schedules, 
operators, and sensors have powered data pipelines for years. As AI agents 
become participants in those pipelines, there's a gap: how does an Airflow task 
hand work off to an autonomous agent and wait for completion?
   
   [GNAP](https://github.com/farol-team/gnap) (Git-Native Agent Protocol) fills 
this gap cleanly. A git repo serves as the coordination board: Airflow tasks 
write to `board/todo/`, agents claim to `board/doing/`, and commit results to 
`board/done/`. Airflow's existing file sensor or a custom GNAP sensor can watch 
for completion.
   
   **Applied to Airflow's DAG model:**
   
   ```python
   # Airflow DAG
   create_gnap_task = PythonOperator(
       task_id='create_gnap_task',
       python_callable=write_to_gnap_board,
       op_args=['board/todo/analyze-dataset-{{ ds }}.md', task_spec]
   )
   
   wait_for_agent = GNAPSensor(  # custom sensor
       task_id='wait_for_analysis',
       gnap_path='board/done/analyze-dataset-{{ ds }}.md',
       timeout=3600,
       poke_interval=30
   )
   
   create_gnap_task >> wait_for_agent >> downstream_processing
   ```
   
   The autonomous agent (LangChain, AutoGPT, custom) picks up the task, 
processes it, and commits to `board/done/`. Airflow resumes the DAG.
   
   This pattern is particularly valuable for Airflow users adopting AI 
incrementally — you keep existing DAG infrastructure while adding autonomous AI 
steps without new message brokers.
   
   Spec: https://github.com/farol-team/gnap


-- 
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

To unsubscribe, e-mail: [email protected]

For queries about this service, please contact Infrastructure at:
[email protected]

Reply via email to