LangChain & LangGraph Introduce Typed Event Streaming for Agent Workflows
Implement LangGraph's new typed event streaming to enable structured UI rendering of messages, tool calls, subagents, and media.
Add LangGraph's new streaming API to your agent runtime, subscribe to desired channels, and build UI components that consume typed projections.
Summary
LangChain and LangGraph have introduced a new streaming foundation that replaces raw token chunks with typed events, enabling structured application streams for modern agents. Each event is labeled with its type—message, tool call, state change, subagent status, or custom channel—and the namespace of the agent tree that produced it, allowing clients to subscribe to specific parts of a graph. The runtime automatically assembles, reorders, and reconnects events, so front‑end frameworks can consume projections such as messages, tool calls, subagent statuses, or custom extensions without parsing raw streams.
Scoped subscriptions let a UI render only the subagents it is interested in, reducing bandwidth and processing overhead. The same protocol works across local runs, remote threads, and React/Vue/Svelte/Angular SDKs, and supports multimodal data like images, audio, and video as separate channels. Custom channels can be defined via transformers to surface domain‑specific streams such as citations or UI descriptions, and ordering metadata enables clients to reconnect and replay from the last seen point. This structured approach turns streaming from a low‑level transport detail into an application API, simplifying the development of dashboards, inspectors, and long‑running production workloads.
Key changes
- Replaces raw token chunks with typed events labeled by work type and agent‑tree namespace.
- Introduces projections (messages, tool calls, state, subagents, custom channels) for clients to consume.
- Supports scoped subscriptions so UIs can render only the subagents they need.
- Same protocol works across local runs, remote threads, and React/Vue/Svelte/Angular SDKs.
- Allows multimodal data (text, images, audio, video) as separate channels.
- Enables custom channels via transformers for domain‑specific streams (e.g., citations, UI descriptions).
- Provides ordering metadata for reconnection and replay from the last seen point.
- Turns streaming into an application API, simplifying dashboards, inspectors, and production workloads.