Two main things worry me about the 'always-on' agent approach:
1. Security & Surface Area: Giving an LLM broad permissions (Email, Calendar, etc.) while it's also scraping arbitrary web content is a prompt injection nightmare. The attack surface is just too wide for production use.
2. Token Economics: Seeing reports of '$300 in 2 days' is a massive red flag. For recurring tasks, there has to be a smarter way than re-processing the entire state every time.
I built Daigest to approach this differently. Instead of an autonomous agent wandering around, it's 'document-centric.' You connect your trusted sources, set a heartbeat, and the AI only processes what's changed to update a structured document. It's less 'magical' than a full agent, but it's predictable, auditable, and won't bankrupt you.
For 'gather and summarize' workflows, a structured document often beats a chat-based agent.