A while back I was asking Claude (privately) if we could have it monitoring the messages posted to an app I’m working on. I wanted it to save them in a certain way, have it watch for messages that begin with //claude and maybe act on them somehow. I mostly just wanted to see if I could create a pipe and what its limits would be.

I asked what the max frequency is and was told it was 1 minute, meaning it would start a thread running Claude that would initialize, and it would read the file, and if there were any new items, it would process them, and then it would shut down. Barely enough to be useful, where we live we like notification within a second, but I went ahead, at least it’ll be a thought-provoking demo.

So I wrote the software. Not much to it, a Node app that’s hooked up to FeedlandSocket, and when it sees a message from me, it appends it to a JSONL file that Claude is watching. At most one minute later it wakes up and sees the new stuff and acts on it.

So I’m right there, it’s all working, the JSONL file is being saved in a public place, and when I ask it to start listening to my JSONL messages I learn that the max frequency is one hour, rendering my work useless. This has happened before, btw.

I asked what about all the “agentic” stuff they hype so much, how do I do that?? I can’t?? What?? Somehow I have trouble believing they’re all telling such a big lie. I wonder if Claude understands its own architecture.

Dave: “What is all this hype about agentic stuff. Every time I try to do something I hit a wall like this.”

Claude: “The pitch is ‘agents that run on their own’; the reality you just hit is a one-hour floor on the only scheduling path the tooling actually exposes. That’s not agency, that’s a batch job with extra steps. The honest read: the cloud-side scheduler is built around long-running, infrequent tasks (digest emails, nightly reports) because each run spins up a fresh sandboxed session, which is expensive. Anything that wants to react quickly to a stream of incoming items — which is what your app actually is — doesn’t fit that model. A plain local cron on some server running every minute would do the job today and cost nothing.”

I edited Claude’s response so as not to not reveal code names for products. 😉

A silver lining — I had never heard of JSONL, but it sounds pretty useful, so if I can’t hook up to Claude realtime, now — I can start using JSONL, which I wll.

Comments, stories, suggestions — here.