By
Yossi Eliaz
AI coding has evolved from single prompts to complex, graph-based workflows. Today’s systems plan, map refactoring tasks across codebases, test code viability, and reduce results into a single pull request. This “Agentic MapReduce” requires an orchestrator built for stateful dependencies and untrusted execution.
Enter Apache Airflow. With recent moves toward sandboxed execution (like PR #68847 and the SandboxExecutor), Airflow is perfectly positioned to orchestrate AI coding workflows.
This session explores modeling AI software engineering as an Airflow DAG. We will cover:
-The AI Coding Graph: Breaking down agentic refactoring into MapReduce DAGs. -The Sandbox Imperative: Why ephemeral containers fail AI agents, and how Airflow’s sandboxed execution secures untrusted model code. -State and Caching: Managing unified caching and persistent state for massive agent-driven build tasks.
Leave with a blueprint for using Airflow as the central nervous system for autonomous software engineering.
Yossi Eliaz
Principal Engineer and Head of DevRel at Incredibuild.com