In the case of HA, the agent is receiving ambiguous natural language requests and converting them into HA MCP calls. I can open the HA app and press buttons, sure, but the agent can reason over time about patterns of water/electricity use, provide natural language feedback, push back on my automation ideas with counterexamples from sensor logs - and all the sensors in the home and on the phones carrying the companion app are rich signal sources for an agent. It unlocks neat tricks like when our babysitter comes over the front door agent can alert her if there's a water leak and walk her through shutting off the water to the house or locate the backup key for her. It can DM me when someone rings the doorbell and carry on a different conversation with STT depending on if it is a person wearing a uniform or holding a clipboard for example thanks to Frigate and some custom vision ingestion with Qwen 3 VL.

Regarding the torrents, that agent keeps track of my ratios on the private trackers and I can give it natural language feedback about what I did or didn't like about a show or movie and it will suggest alternatives, find them on a private tracker, download them, unpack them, and ingest them in the library for me, keeps track of the families like and dislikes and suggests media to watch together, watch private trackers for open invites or interview slots and take the knowledge interviews for me on IRC, deduplicate downloads and hunt for releases with minimum quality standards etc. one of my first tasks for it was replacing my music collection that got damaged when my external hard drive got dropped with no backup and I had to get it extracted in a clean room. Lots of rare lifetime accumulated mp3s gone - mostly recovered now. Searching for replacements would have been a long slog so I automated it.

To your point some of these things could be done by deterministic python scripts and in fact the agents write those to save inference costs later and most of this stuff I could do by siting on my computer an clicking on stuff and reading - the point is that the agents constitute a natural language control surface to trigger all of those actions for me automatically so I can just send a couple chats to a discord bot and it will be done a few minutes later. I'm a busy person I dunno if you can relate. Busy career, got a family to support, parenting obligations, no time to hang out playing video games any more sadly. Building stuff takes time and effort and you have to learn and read about how to do it or pay someone else to do it for you so I was stuck on the easy convenience of windows and Netflix and Amazon/kindle etc. Not anymore! It started with having to install docker desktop to get OpenClaw working and then pretty fast I figured out pretty much everything I wanted to do now that was not work related is better done on Linux and better done self hosted. I would never have gotten over that hump without LLMs.