So I posted about Cowork on web, promising to blog about what I made Hermes agent do, and then… had a whole thing happen with data loss and frantic recovery and backups. That is still ongoing, but in the meantime I can tell you about what Hermes is doing!

Hermes itself is capable of many things, including being run like Claude Code / Codex / etc from the CLI. But I tend to use the Telegram gateway the most, and for unattended actions. Now, unattended actions sound scary, and they are! If you ask it to do “real work”. The things I asked it to do are relatively tame though.

The first thing Hermes is doing is tracking model launches for me — it does Tavily web searches (1000 free credits per month), and fetches Openrouter and major model labs’ pages directly to find out what launched. It then logs what it found to a file, then throws me a summary over Telegram.

The second thing is watching blogs for me — using blogwatcher CLI (I did not know a purpose built tool existed for that?), it watches the Cloudflare Blog, Anthropic’s blog(s) and Simon’s blog (for now). It then filters posts by what it thinks I’ll like (I’m pretty predictable), and sends me summaries and links so I can decide what to read for the day.

Are both strictly necessary for my day? No, not really. But it saves me a few minutes of thinking time per day, and for now I’m happy with that — it seems like what agents are really good for, for now. As they get more capable we might find entirely new capabilities — emergence and all — but for now, this can be done, and for extremely cheaply too (inference pricing is getting quite cheap even for generally pretty capable models!), so I’m happy.

I’ll… blog about what Cowork is doing in a bit. I’m still getting used to how it’s set up, and getting it to work how I want.