Quick Notes: Context Graphs with Apache Airflow^®

AI agents that can't learn from past decisions keep making the same mistakes. Without a full trace of what happened, what was decided by AIs and humans alike, your agents have no way to improve. This quick notes guide walks through a concrete, working implementation of a decision-tracing context graph in Airflow, including orchestrating AI agents using the Airflow AI SDK, human-in-the-loop review via a custom Airflow Slack plugin, and the self-improvement loop that makes agents smarter with every run.

• How to orchestrate AI agents in Airflow using the @task.agent decorator with tool-based access to past decision traces
• How to capture human decisions and reasoning in your Dag using Airflow 3.1's HITLOperator within Slack via a custom Airflow Slack plugin
• How to close the feedback loop by querying the hitlDetails endpoint to feed full decision traces back to your agents