There's a small, deeply human tragedy nobody warns you about when you start using an AI agent: you hit Enter, and then you wait. And wait. The cursor blinks. You tap your fingers. You mentally rehearse your grocery list. Sometimes, in the darkest timeline, you even read the README. It is the universal experience of asking a machine a question and watching it spend eleven seconds remembering how to be a machine.
Well, that intermission just got its curtain pulled. Nous Research shipped Hermes Agent v0.19.0, the Quicksilver Release, and the headline number is a stunner: first-turn time-to-first-token dropped by roughly 80 percent. That's the time between pressing Enter and the first character of the reply appearing, and it got seriously fast.
The Latency That Time Forgot
Time to first token is one of those metrics that quietly ruins your whole day without you ever knowing its name. It's not the same as total output speed. A model can generate a thousand words in a blink, but if it takes a century to warm up and start, every interaction feels like filing paperwork. Agent harnesses made this worse, because there was always a chain of things happening before the model ever saw your prompt.
Quicksilver went after exactly that chain. The result is a snappier feel across the CLI, the desktop app, and the messaging gateway, plus a ~20-PR speed overhaul on the Desktop and TUI streaming so output doesn't crawl. Your own Hermes install gets this the moment you update, no new hardware required. It is the rare performance release that asks nothing of you except to click update.
It's Not Just Speed, Though
If everything else stayed the same, you'd still click update. But Quicksilver also quietly dropped in quality-of-life upgrades that make the raw latency number the least interesting part:
- Smart command approvals are now the default, so Hermes is less trigger-happy and you don't babysit every instruction.
- Bitwarden and 1Password are now integrated as secret sources, a huge win for anyone who never wants to paste a token into a shell history again.
- Terminal and Desktop subscriptions are manageable in-app, and message and subagent delivery are durable across restarts, so nothing you send evaporates when the process dies.
The pragmatic takeaway? The Hermes team spent its last cycle on the opposite of a flashy model launch, building the boring, beautiful infrastructure layer that makes daily use not feel like a root canal.
The Feel-It-To-Believe-It Part
Speed upgrades are easy to fake in a changelog and impossible to fake once you've used it. The 80 percent figure refers to the cold-start first turn specifically, and real users on the ground report the difference is, frankly, rude. You type, you hit Enter, and the answer is already talking like it's been waiting for you all along.
For developers living inside their terminals and anyone running Hermes Agent as a daily driver, Quicksilver is the kind of update you feel in your bones rather than read about. It turns the old waiting-for-the-robot-to-think meme into a nostalgia piece. And honestly? After months of watching cursors spin, that's not just an improvement. That's therapy.
Comments