Why n8n died: In a linear LLM chain (n8n-style), the next step mostly sees the previous steps output. That output is capped. In an agent loop, each tool call is preceded by the growing input transcript (goal, prior calls, results), which can sit in a much larger window and often in cache. n8n-style chains were built for a world where you did not trust the model and the model could not hold much context. You pre-split the job into nodes so each step was small, deterministic, and inspectable That premise aged. Modern models can read on the order of 1M tokens, but they still write far less per turn (often ~64k–128k). A node chain therefore keeps throwing away state: node N+1 inherits whatever node N managed to emit, not the full working set. An agentic loop lets the model keeps one shared context, then appends tool calls and results to it. Each call can be conditioned on the large cached prefix, not on a narrow handoff. Traceability doesnt save the old pattern. “Which node wrote the bad value?” is the same question as “which tool call wrote the bad value?” Both are attribution problems. A graph makes the Step obvious, but it does not make the Fact more trustworthy if the bad payload was already in the stream. Analogy: A node chain is an email pipeline. Writer sends a brief. Artist sees only that brief. Editor sees only what arrived. Nobody shares a desk. An agent is a blackboard. Writer, artist, and editor all read and write on the same board. The board is the state. Tools are hands on that board, not sealed envelopes. --- To be clear, n8n as a product didn't die. The one that is dead is linear-LLM-chain. n8n can wrap an agent and treat workflows as tools. But that isn't how it was commonly used by its users. Most ppl who hyped n8n used it like a linear-LLM-chain When models get more capable and harness like opencode get made, ppl stopped seeing the need for n8n for most of their patterns