What AI agents actually look like in companies: study of 138 real deployments cuts through hype

Researchers analyzed recorded talks from 138 companies discussing how they actually build and run AI agents—systems where a language model takes actions autonomously—to document what architectures, patterns, and technologies practitioners are using in production. This matters because most AI agent discussion is theoretical or promotional; this is documentation of what's working at scale in real industrial systems, which tells you what's actually possible right now versus what's still hype.

Most AI agent discussion comes from vendors selling products or researchers publishing papers on toy problems. This is a direct look at what 138 companies claim they're actually building and deploying—the architectural choices, the repeated patterns, the failure modes they talk about publicly. It's the difference between reading spec sheets and watching how people actually use the tool. If practitioners keep solving the same problems the same way across companies, that's evidence of a converged solution; if they're all doing different things, it means the space is still unsettled and premature claims of standardization are premature.