DAG Agents

A DAG agent structures an AI workflow as a directed acyclic graph: discrete steps (nodes) connected by dependencies (edges), where data flows strictly forward and nothing loops back. Think Airflow, but with LLM calls as the tasks — fetch data, then summarize, then classify, then in parallel draft an email and update a record. The graph is defined before execution starts, either by you at design time or by a model that emits a plan which then runs as a fixed pipeline.

The tradeoff versus a free-running agent loop is autonomy for predictability. A ReAct-style agent decides its next move at every turn, which is powerful but hard to test, bound, and budget. A DAG runs the same shape every time: you know the maximum number of LLM calls, independent branches execute in parallel, failed nodes can be retried in isolation, and every intermediate output is inspectable. When someone asks “make this agent reliable enough for production,” turning the well-understood parts of its behavior into a DAG is often the answer — save the open-ended loop for the steps that genuinely need it.

In practice you’ll reach for a graph framework — LangGraph with only forward edges, or an orchestrator like Airflow, Prefect, or Temporal calling LLMs inside tasks. Each node is a function taking typed input from upstream nodes; fan-out/fan-in handles parallel branches like “summarize each document, then merge.” A common hybrid: a planner LLM generates the DAG dynamically, your executor topologically sorts it, and runs it.