The Quicksilver Command Center + Hermes v0.19.

The Waiting Room Problem.

Imagine hiring the fastest assistant in the world.

Then making them sit in a waiting room before every single task.

That's what running AI agents felt like before this week.

You ask a question — and stare at a blank screen for four, five seconds before anything moves.

You hand an agent a big job, close the laptop, and come back to find the job died with the app.

You run one bot for everything, so your SEO agent answers with your content agent's brain.

And every risky command needs you standing there, clicking "approve" like a hall monitor.

Each wait is small.

Added up, it's the reason most people poke at AI agents instead of running a business on them.

The Quicksilver Command Center breaks that cycle for good.

Same agent. Liquid-metal reflexes.

Hermes is an open-source AI agent from Nous Research.

It lives on your computer, talks to any AI model you point it at, and does real work — files, browsers, messages, schedules.

v0.19 is called the Quicksilver Release, and the name is literal.

The team rebuilt how fast it wakes up, how fast it answers, and how safely it can run many agents at once.

Three ideas cover the whole release:

1. Answer instantly. The time from pressing enter to seeing the first word dropped by about 80%. Cold start went from 4.3 seconds to 0.9.

2. Never lose work. Background jobs and final replies are now written into a ledger. If the app or gateway crashes mid-job, it finishes and delivers when it comes back.

3. Command a fleet, not a bot. One gateway can now route different channels to completely different agent profiles — each with its own model, memory, skills and secrets.

One message's journey through v0.19.

So when someone asks "but what actually changed?" — the honest one-liner: the same Hermes agent now answers in a second, survives crashes, and runs many specialist agents behind one door.

The official sources. Read it yourself.

Replies that land 80% faster.

Let me start with the speed, because it changes how the agent feels.

Before this release, asking Hermes anything meant a little wait — the app warming up, the model spinning, a blank line.

Now the cold start is 0.9 seconds, and reasoning models stream their thinking live, so something useful is always on screen.

You ask in your Telegram on the way to the gym. The answer is forming before you've pocketed the phone.

The desktop app got the same treatment — 14× less CPU on markdown, no more freezes on big diffs, instant session switching.

Background agents you can watch working — and that survive a crash.

Next big one. This is my favourite in the whole release.

Hermes could already delegate work to background subagents — hand a research job off, keep chatting.

The problem was trust. The work happened in a black box, and if the process died, the job died with it.

v0.19 fixes both halves.

Live subagent transcripts mean you can open any background agent and watch its actual conversation, mid-job — like glancing over an employee's shoulder.

Durable delegation means completions are written to an ownership-checked ledger. Restart the app, reboot the machine — the finished work is still there, delivered.

You hand your agent "research these 20 leads" at lunch. Your laptop dies at 2pm. The results are waiting at 3.

And the same durability now covers the gateway: final replies are recorded around the platform send, and redelivered on the next boot if anything crashed in between.

Smart approvals: a second brain checks the risky stuff.

If you have ever left an agent running and felt a little knot in your stomach — this section is for you.

Before, you had two bad choices: approve every flagged command by hand, or switch on full-auto and hope.

v0.19 adds a third way, on by default: when a command gets flagged, a separate AI independently assesses it before it runs.

On top of that, your own deny rules now block commands even in full-auto mode. Full-auto no longer means no brakes.

And when you refuse something, /deny too aggressive — draft it as an email instead passes your reason back, so the agent course-corrects instead of guessing.

One gateway, many agents — routed by channel.

This one is the sleeper hit of the release.

A single Hermes gateway can now route messages to different profiles depending on the guild, channel or thread they arrive in.

Each profile is a fully separate agent: its own config, its own skills, its own memory, its own secrets.

Your Discord #seo channel talks to the SEO specialist. #content talks to the writer. Your private thread talks to the agent that knows your numbers.

One server process. A whole staff.

I run thirteen Hermes profiles — one per model and job. Before, they were separate doors. Now they can sit behind one switchboard.

Here's what that looks like on my machine — my Hermes hub inside the Agent OS, with the profile row along the top:

What you're looking at: my real Hermes hub inside the Agent OS. The pill row is the fleet — julian, grok-4-5, gpt56, hermes-cloud, jarvis, local and more. Every one is a separate agent with its own brain, and v0.19's multiplex gateway can route channels straight to them.

The smaller upgrades that add up.

Quick tour of the rest — the things you'll actually touch week one.

Reasoning dials. New max and ultra effort levels, a session /reasoning command, and /model --once for a one-turn model switch that reverts by itself. Ask one hard question on the big model, drop back to cheap automatically.

Stacked skills. Chain slash-skills in one line — /research /write-report quarterly numbers — and they dispatch as a single agent turn.

Real secrets management. Bitwarden and 1Password now plug in as secret sources, multiple vaults at once, with clear precedence. API keys stop living in text files.

Sessions you can hand to anyone. Export any session to Markdown or HTML — and a --redact flag scrubs secrets before it leaves your machine.

Billing without the dashboard. /subscription shows your plan and remaining allowance; /topup upgrades or downgrades right from chat.

A panic button. --safe-mode boots Hermes with every customisation off — the "is it my config or the app?" question, answered in one flag.

Nicer buttons everywhere. Telegram, Discord and Matrix get native choice pickers; WhatsApp clarify prompts render as actual polls.

What actually changes in your day.

The Quicksilver Command Center.

Here's how I run all of this as one system — five layers, from the metal up.

Install Hermes and you get the parts.

Wire the five layers together — that's the Quicksilver Command Center, and it's exactly what the Agent OS ships pre-built.

What's actually holding you back.

Step by step — five minutes, verified.

Every command here ran on my machine before it went in this guide.

Step 1: Update Hermes

hermes update

That's the entire migration — it's the exact command Nous posted under the announcement.

Step 2: Confirm you're on Quicksilver

hermes --version

You'll see: Hermes Agent v0.19.0 (2026.7.20). That's the confirmation you're on the new core.

Step 3: Check your default brain

hermes config get model.default

Mine answers gpt-5.6-sol. The new config subcommands mean you never dig through YAML to check a setting again.

Step 4: Feel the speed

hermes chat -q "What changed in v0.19?" -Q

Watch how fast the first word lands. That's the 0.9s core plus live-streamed thinking.

Step 5: If anything acts weird

hermes chat --safe-mode

Boots with every customisation off — instantly tells you whether a problem is your config or the app.

And this is where it all lands in my world — the Hermes Oracle inside the Agent OS, a Hermes agent that wakes at 6am, scans the AI news, and briefs me:

What you're looking at: the Hermes Oracle tab in my Agent OS — live signals, last consulted timestamp, and a real news card it pulled and summarised on its own. This is the kind of always-on surface the v0.19 speed and durability upgrades make painless to run.

Yes — and here's the honest version of why.

This isn't a "new model" release you can skip.

It's the release that makes the agent you already have feel instant, and makes leaving it alone safe.

If you run Hermes at all: hermes update today.

If you've never run it: this is the best week there's ever been to start — the speed problem that made beginners bounce is gone.

The people who figure out AI agents now, while the tools are evolving this fast, are going to be way ahead when everything settles. Every update you learn compounds.

The recap.

I'll see you in the next one.