My Claude Code agents spent 2,300 hours waiting for things to happen. The model ran 35 times.
retained text
I run a bunch of Claude Code agents for personal stuff, and a lot of what I ask them is "tell me when X happens". When the DHL email comes. When I get home, remind me to take the trash out. When the bank texts about a payment.
The obvious way is a scheduled job that wakes the agent every 15 minutes to go check. Almost every one of those turns would end with "nothing yet", and each one reloads the whole context to say it.
So the agent only writes the watch when I ask (which sensor, what to match, what to do when it fires) and then a plain script owns it. One job every minute, no model in it, and it only samples watches that are due.
Gmail is a search query with after:<when the watch was made>, all of them batched into one call. Home is my phone showing up on the home wifi, found with ARP so a sleeping phone still counts. Location is my shared Maps position, and the next check is half the time I could possibly get there at 100 km/h, so 40 km out it checks every ~12 min and the last km every minute. SMS gets read off Messages Web.
When one fires, a plain reminder is fixed text sent to the chat, no model. Only a watch with a task ("summarize the reply") wakes the agent.
17 days, plain time reminders not counted: 30 event watches, 2,316 watch-hours open, 10 fires. 25 model calls, all of them one SMS watch where the sender also sends marketing, so Sonnet decides if a text is a real notice. Plus the 10 fires. Polling every 15 min would have been around 9,000 agent turns for the same watches.
Yes it's still cron underneath. But the cron runs code. Matching a sender or a phone on the network costs nothing, I only need the model to set it up and to act once something happened.
Downside is every sensor is code I have to keep alive. Right now the Maps share link stopped returning a position, so location watches are dead until I send a new link. At least it told me after 3 failed checks instead of going quiet.
Anyone else pulled polling out of the agent? What sensors did you end up with?document fetched 2026-11-03 00:22 · canonical https://reddit.com/r/ClaudeAI/comments/1xlxw0q/my_claude_code_agents_spent_2300_hours_waiting