What it is

Linea is a canvas for building AI workflows: connect an LLM call, a retrieval step, a tool call like Stripe or Slack, wire them together, and watch data move between them instead of debugging a prompt chain you can't see inside. It's live at getlinea.app, still growing.

The hard part wasn't the AI

The LLM calls are commodity. The actual engineering sits underneath, in a state machine that tracks every node through idle, running, streaming, and complete or error, because a streaming LLM node isn't "running" and then "done", it's emitting tokens the whole time, and whatever's downstream needs to know whether a given token is a partial result to buffer or something it can act on immediately. That state also has to stay fast. Early on, streaming token state lived in React state, and the canvas re-rendered thirty-plus times a second during generation, stuttering visibly on anything longer than a couple of nodes. Moving that state into a ref-based store that components subscribe to directly, bypassing React's render cycle entirely, fixed it: the canvas now only re-renders on structural changes, like a node's status or an edge being added, never on individual tokens. Memory across a multi-step workflow follows the same logic. Each node declares an explicit dependency on another node's output, wired as an edge on the graph, so the execution engine only ever passes a step what it actually needs, no implicit global context. pgvector backs long-term retrieval underneath that, but the graph-level wiring is what keeps the whole thing debuggable instead of another pile of state nobody can see.

Stack

Next.js · NestJS · LangGraph.js · PostgreSQL · pgvector · Redis · BullMQ · Clerk

Where it stands

Deploying a custom agent went from weeks of bespoke integration work to something closer to wiring a diagram together, roughly 70% faster to get one live, and inspectable by whoever needs to trust it, not just whoever wrote the prompt. More on how the execution engine works in Building Linea.