Most LLM agents running in the wild skip the safety nets.
Researchers analyzed more than 6,000 publicly available workflows on n8n, a low-code automation platform, to produce what they describe as the first large-scale empirical study of how non-expert users actually wire up LLM agents in practice. The findings complicate the tidy picture that AI vendors tend to sell. LLMs in these workflows aren't sitting at the end of a simple prompt-and-response pipe — they're embedded inside broader automation structures that pull in external APIs, communication services, storage systems, and occasional human review steps. The study examined task distribution, structural patterns, tool use, reliability mechanisms, and how much autonomous decision-making each workflow delegates to the model.
The uncomfortable finding is the gap between deployment and engineering discipline. Explicit reliability mechanisms — structured fallback paths, repair loops, failure-specific alerts, human approval gates — are relatively uncommon across the sample. That matters because these workflows aren't toys: they're connecting LLMs to live external services where a bad output or a silent failure can propagate downstream before anyone notices. The researchers frame this as a gap between how fast LLM agents are being adopted and how slowly the surrounding infrastructure for safety and governance is catching up.
This tracks with a pattern visible across the broader automation industry: tooling lowers the barrier to build, but not the barrier to build well. n8n is not alone here — similar dynamics almost certainly show up on Zapier, Make, and every other platform racing to bolt an AI layer onto workflow automation.