Two agents, one workflow

I wired up a Hermes agent on DeepSeek V4 Flash this week and handed it two jobs I'd been doing by hand: keeping a journal and writing a newsletter. The setup took an afternoon. The payoff was immediate.

The first agent reads my day — commits, scattered notes, half-finished thoughts — and turns it into a journal entry I'd actually read back later. The second takes the same raw material and drafts a newsletter in my voice, not the model's.

What surprised me wasn't the writing. It was how little babysitting it needed once the prompts were tight. Flash is fast and cheap enough that I stopped thinking about cost and started thinking about the workflow.

This is the pattern I keep coming back to: AI in the loop, not in the driver's seat. It removes the bottleneck — the blank page, the context-switch — without pretending to be me.

More on the actual prompt structure and guardrails soon.