The Herald Upgrade™ — Hermes v0.20, tested live.
Hermes Agent just shipped its biggest free update ever — and your AI can finally hold a real conversation.
You talk, it answers as it thinks, and you can cut it off mid-sentence just by speaking — like a phone call, not a voicemail.
Say your own wake word from across the room, and your computer starts listening.
It backs research with sources you can actually check, pings your other systems the moment work finishes, and even talks to other agents.
I updated my whole Agent OS to it and put the big claims to the test — the real numbers are below.
And one new command moves your entire Claude Code setup into Hermes in a single line — stick with me to the end.
"The Herald Release. Hermes is the herald of the gods, and this release makes him one in earnest: he speaks… he carries word to other agents… he announces events to your systems… and he cites his sources."
— Nous Research, official v0.20.0 release notes, August 3, 2026
The Friction Problem — powerful agents that are painful to steer.
Here's what actually slows people down with AI agents. It isn't intelligence. It's friction.
The agent heads the wrong way, and your only move is to stop it and re-explain from zero.
Voice sounds like the future until you try it — speak, wait, listen to a monologue, repeat.
Research comes back confident but unsourced, so you re-check everything yourself.
Long jobs die at an arbitrary iteration wall. Integrations mean polling loops. Tools fail and the model guesses why.
Each one is small. Together they're the tax you pay on every single session.
v0.20 goes after that tax directly. That's why I call it the Herald Upgrade™.
"My agent already works. Why chase an update?"
Because the friction is invisible until it's gone. You've adapted to restarting turns and re-checking sources — the update removes the adaptations you forgot you were making.
What the Herald Release actually is.
Hermes Agent is the open-source AI agent from Nous Research — the one my whole Agent OS runs on.
Version 0.20 is one free update with four big ideas inside it.
First, Hermes learned to talk properly. It speaks while it thinks, and you can interrupt it by just talking.
Second, it learned to prove things. Research now comes back with citations matched to the real page text, plus a fact-checking mode.
Third, it learned to reach out. It can push signed events to your other systems, and speak the Agent-to-Agent protocol with other AI agents.
Fourth, everything got smoother — steering, tools, memory, speed.
A voice conversation with Hermes, from wake word to answer.
You say your wake phrase
"Hey Hermes" — or any phrase you pick. Detection runs on your device, so no audio leaves your machine while it waits.
You ask in plain words
Speech-to-text transcribes you — in the right language, with the new unified language handling.
It answers while it thinks
Speech streams clause by clause as the response generates. No waiting for the full reply before you hear anything.
You interrupt by talking
It stops, listens, and the model is told you cut in — so the next answer accounts for it.
You say "stop" and it ends
Hands-free the whole way — CLI, desktop, or a WhatsApp voice note. Same conversation, every surface.
Three doubts, answered before the update.
What changes in a normal working day.
Type, wait, restart.
- Voice = voicemail: speak, wait, listen to one long file.
- Wrong direction? Stop the turn and re-explain from zero.
- Research answers you re-verified by hand.
- Integrations meant polling Hermes from outside.
- Long jobs died at the 90-iteration wall.
- Failed tools made the model guess what went wrong.
Talk, steer, trust.
- Speaks as it thinks — interrupt by just talking.
- Correct it mid-turn; work in flight is preserved.
- Citations matched to real page text, fact-check included.
- Hermes pushes signed events to your systems — no polling.
- Ceiling raised 90 → 500 for long autonomous runs.
- Tools diagnose and recover from their own failures.
What I measured on my own Mac.
This whole Agent OS runs on Hermes, so I updated it the minute the release landed and checked the claims.
"Is a 3,650-commit update safe on a system I depend on?"
That was my test. My OS runs Hermes in six different places, and the one-command update broke none of them. The release notes also document every change with its pull request, so nothing lands unexplained.
Every headline feature, one by one.
Talk to it like a person
Hermes speaks clause-by-clause as it thinks. Interrupt mid-sentence by just talking — it stops, listens, and knows you cut in. Example: ask for a summary, hear the first line, cut in with "just the numbers" — it pivots without restarting.
Hands-free from across the room
Pick any phrase and detection runs on your device — no audio leaves your machine while it waits. Different wake words can reach different profiles. Example: "hey Hermes, how did the overnight build go?" from the kitchen.
Research with receipts
Every claim gets a verifiable source, quotes are matched against the actual page text, and fact-check mode grades any document you hand it. Example: paste a competitor's claims page and get back what's true, false, and unverifiable.
Hermes calls YOUR systems
Signed lifecycle events push to any HTTP endpoint you register — with signatures your receiver can verify. Example: your dashboard lights up the moment an overnight agent run completes, no polling loop.
It speaks to other agents
The Agent-to-Agent protocol landed — Hermes can discover, talk to, and be driven by other A2A-compatible agents. Example: a planning agent on another stack hands Hermes a task and gets the result back over the standard wire.
The CLI power wave
Run a shell command instantly with ! — no model turn spent. /init writes your AGENTS.md, /diff shows changes, /context reveals what fills your window. Example: !git status mid-conversation, zero tokens.
Steer without restarting
Type a correction while it works — the turn redirects, in-flight work is preserved, and the agent course-corrects. Example: "actually, make it dark mode" lands while it's still building, and nothing done so far is thrown away.
Fewer wasted turns
Truncated output spills to a readable file, patches diagnose their own mismatches, empty searches probe for near-misses. Example: a huge build log no longer vanishes — the agent reads the spilled file and carries on.
A workbench, not a chat
Artifacts render with sandboxed live preview beside the chat, a plugin SDK landed with Kanban first, and a global hotkey captures a thought into any session. Example: generated HTML runs safely in the right rail while you keep chatting.
Six ways the Herald changes your week.
Drive work by voice
Kick off builds and check agents from across the room — no keyboard, no screen.
Client-ready research
Reports where every claim links to checkable evidence — send them without re-verifying.
Systems that react
Webhooks ping your dashboard or phone the moment an overnight job lands.
Mixed agent fleets
A2A lets Hermes join multi-agent systems built on other stacks — one protocol.
Longer autonomous runs
The 500-iteration ceiling plus self-fixing tools means big jobs finish instead of stalling.
One-line migration
import-agent moves your Claude Code setup — skills, config, memory — into Hermes.
If you want this in one place.
Everything in this guide runs inside my Agent OS — Hermes wired into every agent, studio and dashboard, with the voice, webhooks and citations you just read about. Inside the AI Profit Boardroom you get it pre-configured, plus the people to help when something doesn't work first try.
- The Herald Upgrade™ — Hermes v0.20 wired through the whole OS
- The complete Agent OS — every agent and studio in one dashboard
- 5 live coaching calls a week — get unblocked by people on your exact setup
- 1,000+ prebuilt agents and the automation playbooks
- 4,000+ members in 38 countries — someone's online whenever you're stuck
Real people are running this stack right now.
Members inside the Boardroom run Hermes-powered systems across agencies, ecom stores and content businesses.
Members post their wins in a 158-page doc — read it here →
What this costs, straight up.
The update is free. Hermes Agent is open source, and v0.20 installs with one command over whatever you're running.
Model usage is whatever your providers charge — same as before. My OS routes most Hermes work through free models, and that setup kept working on v0.20.
The honest belief shift here is small but real: an update this size used to mean a broken week — now it's one command and a version check.
Follow the steps below and you're on the Herald Release tonight. Or take the shortcut — the Boardroom has the whole OS pre-wired with live help.
Eight steps into the Herald Release.
Step 1: Update
One command, a few minutes:
hermes updateStep 2: Confirm the build
Run hermes --version — you want v0.20.0 (2026.8.3). Mine came up first try.
Step 3: Have your first real conversation
Start voice mode and interrupt it mid-answer on purpose. Feel the difference — that's barge-in.
Step 4: Set your wake word
Pick your phrase and enable on-device wake detection. Your terminal now listens from across the room, offline.
Step 5: Initialize your project
In your repo, run /init — Hermes scans the project and writes your AGENTS.md for you.
Step 6: Try grounded research
Ask a research question with the grounded-citations skill and click the evidence links it returns.
Step 7: Mine your approvals
Let it learn what you always allow:
hermes approvals suggestStep 8: Migrating from Claude Code?
Bring your whole setup across in one line:
hermes import-agent claude-codeThirty days with the Herald.
Week 1 — The voice
Days 1–7Goal: updated, wake word set, first tasks driven entirely by conversation.
Week 2 — The trust
Days 8–14Goal: grounded citations on every research task; first fact-check run on your own content.
Week 3 — The wiring
Days 15–21Goal: webhooks pushing events into your dashboard or phone; approvals allowlist mined and applied.
Week 4 — The fleet
Days 22–30Goal: long autonomous runs on the 500 ceiling, and A2A connecting Hermes to your other agents.
Everything you just learned, in one glance.
You update free
One command takes any install to v0.20 — mine upgraded a whole OS clean.
You talk naturally
It answers while it thinks, and you interrupt by just talking.
You go hands-free
Your own wake word, detected on-device, "stop" to end.
You trust the research
Citations matched to real page text, fact-check verdicts included.
You wire your systems
Signed webhooks push events — your dashboards react without polling.
You steer live
Mid-turn corrections keep in-flight work instead of restarting.
You run longer
Self-fixing tools and a 500-iteration ceiling finish the big jobs.
You migrate in one line
import-agent brings your Claude Code setup with you.
Update tonight. Or take the shortcut.
One command puts you on the Herald Release. And if you want the whole Hermes-powered Agent OS around it — pre-wired, with people to call — that's the Boardroom.
- The Herald Upgrade™ + the full Agent OS, ready to run
- 5 live coaching calls a week + a community already on v0.20
- 258 documented member wins — collected in the 158-page doc inside
