Your AI agent should not make you wait. The hermes agent quicksilver release — officially Hermes Agent V0.19 — is built around that one idea. There are no flashy new toys here: 2,245 commits landed in a single release, and almost every one of them is aimed at speed and reliability. Julian Goldie updated the day it shipped, and the difference is obvious within the first minute of using it.
Here is what Quicksilver actually changes, how the speed feels in daily use, and how to update in one click.
What Is Hermes Agent Quicksilver?
Quick context first. Hermes is an open-source AI agent that lives on your computer and talks to any model you point it at. In Julian Goldie's view, it is more powerful and more customisable than any paid agent he knows of — the harness itself is free, and it can run free models on top. It also sits at the centre of his Agent OS, the system he uses to run his business on AI agents.
Quicksilver is the codename for Hermes Agent V0.19. Where earlier releases chased headline features, this one goes after the unglamorous fundamentals: how fast the agent responds, and whether it finishes what it started. That focus shows up everywhere — all 2,245 commits of it.
The Speed Numbers: 4.3 Seconds Down to 0.9
The headline stats:
- Replies land around 80% faster than on V0.18.
- Cold start dropped from 4.3 seconds to 0.9 seconds.
- The first word appears in under a second.
The practical difference is bigger than the numbers suggest. On the old version, you sent a message and stared at a blank screen while the agent woke up. On Quicksilver, you watch the model thinking in real time — the reply starts forming almost as soon as you hit enter.
For context, Julian Goldie runs every agent and model he uses through Goldie Bench, his own hands-on testing scoreboard. In that testing, slow first replies are one of the most common frustrations — Grok 4.5, for example, consistently felt slow to reply when he tested it. On his bench, Quicksilver's under-a-second first word makes Hermes feel snappier than anything else he has run, paid or free. That is his experience, not a lab measurement — but it was convincing enough to update on day one.
Fleet Routing: One Gateway, Many Specialist Agents
This is the biggest structural change in Quicksilver. One gateway can now route different channels to completely different agent profiles.
Each profile is a fully separate agent with its own model, its own memory, its own skills and its own secrets. In practice, that means you can run one profile for SEO, one for writing and one for ops — and each behaves like a genuine specialist instead of one generalist trying to answer everything with the same brain.
Auto-Delegation With the Delegate Tool
You do not have to manage each profile by hand. The main Hermes agent can auto-delegate to sub-agent profiles using the delegate tool. You talk to one agent, and it quietly hands work to the right specialist behind the scenes. One conversation, a whole fleet doing the work.
Two New Free Models: HY3 and Laguna S 2.1
Quicksilver also adds two new free models you can use with Hermes: HY3 and Laguna S 2.1. A practical setup looks like this: give different profiles different free models, keep a local model in the mix, and if one provider gets rate-limited, the main agent automatically routes around it. The fleet keeps working even when one model does not.
📺 Watch: Hermes Agent V0.19 Just Changed AI Agents Forever!
Smart Approvals: Autonomy With Brakes You Control
Approvals used to force a bad trade: either hand-approve every flagged command one by one, or switch to full auto and hope nothing dangerous slips through. Quicksilver adds a third option — and makes it the default.
- Manual mode — you approve every flagged command yourself.
- Full auto — everything runs without asking.
- Smart approvals (the new default) — when a command gets flagged, a second AI independently assesses it before it runs.
On top of that, you can set your own deny rules — and they hold even in full auto. Julian Goldie sums it up as the "speed of autonomy with brakes you control". You get far fewer interruptions without giving up the final say over what your machine is allowed to do.
Reliability: Nothing Dies Silently Any More
The other half of Quicksilver is trust. Background jobs and final replies are now written to a database. If the app or the gateway crashes mid-job, the agent still finishes the work and delivers the result when it comes back.
On V0.18, a crash meant background jobs broke silently — you found out hours later that nothing had happened. Now the job survives the crash. For anyone running scheduled or overnight work through their agent, this is arguably the most important change in the entire release.
📺 Watch: Hermes Agent + Buzz is INSANE!
Old Way vs New Way: V0.18 vs Quicksilver
| Area | Old way (V0.18) | New way (Quicksilver) |
|---|---|---|
| Cold start | 4.3 seconds | 0.9 seconds |
| First reply | Staring at a blank screen | First word in under a second; you watch it think live |
| Crashes mid-job | Background jobs broke silently | Jobs and replies stored in a database, delivered on recovery |
| Channels | One agent answering every channel with the same brain | Fleet routing to fully separate profiles per channel |
| Risky commands | Hand-approve one by one, or full auto with no brakes | A second AI assesses flagged commands; your deny rules always hold |
| API keys | Sitting in plain-text config | Real secrets management |
| Sharing sessions | Copy-paste and hope secrets stayed secret | Export to markdown or HTML with a redact flag that scrubs secrets first |
The Smaller Upgrades Worth Knowing
Beyond the headlines, Quicksilver ships a stack of quality-of-life improvements:
- Reasoning dials and new stat skills.
- Real secrets management — no more API keys living in plain-text config files.
- Session export to markdown or HTML, with a redact flag that scrubs secrets before anything leaves your machine.
- A /subscription command for handling billing.
- Safe mode — boot with every customisation switched off if your config ever breaks.
- Nicer buttons on Telegram, Discord and Matrix.
Small things on their own — but stacked together they remove a dozen daily frictions.
📺 Watch: Hermes Agent: New Voice Activation Update is INSANE!
Why Quicksilver Matters Inside Agent OS
Because Hermes sits at the centre of Julian Goldie's Agent OS, every workflow built on top of it inherits the speed automatically:
- Kanban delegation is faster — tasks hand off to agents with far less lag.
- Memory retrieval is faster — his "memory galaxy" pulls context back more quickly.
- Custom workflows all benefit — Hermes Apollo (his voice agent), Hermes Oracle (news) and Hermes Astros (competitor monitoring) all run on the same engine, so all of them got quicker overnight.
- Fewer interruptions — the second-AI approval gate means far fewer approve-this pop-ups breaking your flow.
A speed upgrade to the engine is a speed upgrade to everything attached to it.
If you want the fastest Hermes agent setup making you money, check out the AI Profit Boardroom — the full Agent OS install, fleet routing profiles and free-model stack are already set up inside. → Get the Quicksilver fleet working for you
How to Update to Quicksilver
Two options, both quick:
- Click Update Hermes in the Agent OS manage section, or
- Run hermes update from the terminal.
And if an update ever breaks your configuration, the new safe mode boots Hermes with every customisation switched off so you can fix things calmly.
FAQ
What is Hermes Agent Quicksilver?
Quicksilver is the codename for Hermes Agent V0.19 — a release of the open-source Hermes agent focused on speed and reliability. Across 2,245 commits, it delivers faster replies, crash-proof background jobs, fleet routing and smarter approvals.
How much faster is Quicksilver than the old version?
Replies land around 80% faster, cold start has dropped from 4.3 seconds to 0.9 seconds, and the first word of a reply appears in under a second.
Is Hermes free to use?
Yes. Hermes is open source and the harness costs nothing. With Quicksilver you can also run the free HY3 and Laguna S 2.1 models, plus a local model, so a complete working setup does not have to cost anything in model fees either.
What happens if Hermes crashes mid-task?
Background jobs and final replies are written to a database. If the app or gateway goes down mid-job, the agent still finishes the work and delivers the result when it comes back — no more silently broken jobs.
Is full auto mode safe now?
Much safer than before. The new default has a second AI independently assess any flagged command before it runs, and your own deny rules hold even in full auto. You choose the level of autonomy; the brakes stay yours either way.
The Bottom Line
The hermes agent quicksilver release is the rare update that improves everything without asking you to learn anything new. Replies around 80% faster. Cold starts down from 4.3 seconds to 0.9. Jobs that survive crashes. A fleet of specialist agent profiles behind one gateway, two new free models in HY3 and Laguna S 2.1, and approvals that finally balance speed with control.
Julian Goldie updated the day it shipped, and on his own testing it made Hermes feel faster than any paid agent he has tried. One click on Update Hermes — or one hermes update in the terminal — and every agent workflow you run gets quicker at once.
