Hey, I have a technical question regarding Claygent. Multiple people I talked to claim that Claygent can run 5k rows in less than a minute. That's an insane number. Does anyone understand (technically) how Clay does it? What's the concurrency level for the agent, what does the unit economics look like and so on.
Sandro T. I would say you're probably getting bad information. A single claygent typically runs in 2-5 minutes, depending on the model and the complexity of the task. Running 5k claygents would take more like 30 minutes or more typically - they run in large batches concurrently, but I dont know what the batch size is exactly.
However Clay does have a number of API-based enrichments that can process larger amounts of data more quickly (essentially just a database query) as well as formula enrichments which can handle a lot of basic logic and transforms for free and instantly.
Sandro T. Personally, I don't think Claygent is actually processing 5,000 rows in under a minute if each row requires genuine AI research. 5,000 rows in 60 seconds is about 83 rows per second. Even with heavy parallelization, that seems unlikely if every row involves web searches, reasoning, and LLM calls. My guess is that when people quote numbers like that, a lot of those rows are either cached, using pre-existing enrichment providers, or running much simpler tasks than full Claygent research.
I thought so. The number would be crazy, even if you put it at 500 concurrency and theoretically reduce the scraping part to 0 seconds, I'm not sure you'll get all the LLM responses in 1 minute. It's probably an overstatement
