Author(s): Praveen Kumar

Originally published on Towards AI.

LlamaIndex Workflows Is Now a Standalone Package. Its Typed State Is What Makes That Matter.

If you learned to build LlamaIndex agents by writing from llama_index.core.workflow import Workflow, that import is now a compatibility shim. The orchestration engine underneath it has moved out on its own. As of the Workflows 1.0 announcement, the code that routes events between your agent's steps lives in a separate package — pip install llama-index-workflows — that imports as plain workflows and does not depend on llama_index at all.

The article explains that the big shift isn’t just the new package import path, but that Workflows now uses typed Pydantic state in a generic Context[MODEL_T], enabling runs to be frozen to JSON mid-execution and later restored without rerunning steps. It demonstrates a deterministic triage workflow that branches via union return types, loops by returning a StartEvent, and streams progress through events that are emitted but not consumed by steps. After showing the workflow run output, it proves the “checkpoint + resume” capability by serializing the full context to a small JSON blob and rebuilding it via Context.from_dict(), yielding validated state on restore. Finally, it clarifies where the standalone package fits (control flow, typed state, streaming, serialization; optional observability) and where it doesn’t (not a batteries-included agent with built-in RAG/tool-calling), and concludes with key takeaways about the engine split, the meaning of “1.0,” and why typed state is the core reason to care.

Read the full blog for free on Medium.

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI