Talking Agents Have Arrived: Hermes Agent Finally Speaks
Hermes v2026.8.3 update explained in plain English — talking agents, agent-to-agent messaging, and verifiable citations.
Version Information
| Item | Details |
|---|---|
| Version | v0.20.0 (Tag: v2026.8.3) |
| Release Date | August 3, 2026 |
| Release Codename | The Herald Release |
| One-Line Summary | Agents can finally “speak” out loud, relay messages to other agents, proactively report to your systems, and back up everything they say with sources. |
Plain-English Breakdown
🎙️ Chat with AI like a phone call, not a walkie-talkie
Before, talking to Hermes felt like using a walkie-talkie: you’d say your piece, it would pause for a long moment, then let out a whole recorded monologue with a “beep—”. Not anymore. Now it thinks and talks at the same time, delivering responses sentence by sentence, just like a real conversation. Want to jump in mid-sentence? Just speak up — it’ll stop immediately and listen, and it knows you interrupted, so it won’t keep rambling on. It’s like upgrading from a voicemail inbox to a live phone call — no more staring at the ceiling waiting for it to finish its script.
🗣️ Just say “Hey, Hermes” to wake it up — fully offline
You can set a custom wake word, like “Hey Hermes” or literally anything you like. The moment you say it, it starts listening. The key part? Wake word detection happens entirely on your device — no audio is sent to the cloud. That means everything you say before it “wakes up” never leaves your computer. Privacy maxed out. Plus, different devices can have different wake words — say, “Hey Hermes” on your living room PC and “Hey Butler” on your bedroom tablet, with zero cross-talk.
📣 Agents can now “pass messages” to each other — no more working in silos
Previously, if you asked an agent to do a task, it would finish and stop there — no way to hand off the result to another agent. This update introduces the A2A v1.0 protocol — agents can now message each other and hand off tasks like coworkers. For example, you can have Agent A summarize meeting notes, and it’ll directly “pass” the summary to Agent B, who then generates the to-do list. The whole flow works like different departments collaborating in a company, instead of every agent having to ask you from scratch.
🔔 It proactively “knocks on your door” with updates — no more waiting for you to check
Before, agents would go silent after finishing a task, and you’d have to go check yourself. Now it supports signed outbound webhooks — in plain terms, when it finishes something, it actively “knocks” to notify your other systems (like sending a message to your WeCom, DingTalk, or a custom system). And that notification comes with a digital signature, so you know it’s genuinely from the agent and not something forged. It’s like a delivery driver texting you after a signed delivery, instead of making you refresh the tracking page yourself.
📚 Everything it says is backed by sources — no more making things up
This one’s the most reassuring: Hermes now cites its sources when doing research or answering questions. Every key claim comes with clickable links to the origin, and it can even do fact-checking. Before, if you asked “What was Company X’s revenue last year?”, it might give you a number with no idea where it came from. Now it’ll tell you: “This data comes from this report — here’s the link.” It’s like writing a paper with references, not just pulling numbers out of thin air.
🛠️ When tools fail, it picks itself up — no more manual retries
Previously, if a tool (like a weather plugin) temporarily errored out, Hermes would stop and ask you “What should I do?” or give you a vague error message. Now it’s learned self-healing — when a tool fails, it automatically retries with a different approach or gracefully falls back to a backup plan, instead of dumping the mess on you. It’s like a food delivery rider who, when the restaurant is closed, just picks another place instead of calling you to ask “should I switch?”
💻 Desktop app and CLI both got serious upgrades
The desktop version now supports multiple windows, live preview (e.g., seeing charts render as you write the code), and an open plugin SDK so you can build your own extensions. The CLI got a bunch of handy shortcuts: press ! to drop into system shell mode, /init to quickly scaffold a project, /diff to compare file changes, /context to inspect the current context, and /focus to zero in on a specific task. For tinkerers, this roughly doubles your workflow speed.
Summary
The biggest takeaway from this release: Hermes has evolved from an “obedient tool” into a “communicative partner.” It’s no longer a bot that only answers when asked — it proactively speaks up, reports progress, collaborates with other agents, and even cites its sources like a reliable colleague. Whether it’s the fluidity of voice interaction, the collaboration between agents, or the trustworthiness of the information, everything has taken a major leap forward. If you previously thought AI agents were a bit “slow on the uptake,” this release is worth a second try — it’s finally starting to get how humans talk.
📎 Official Release Notes: https://github.com/NousResearch/hermes-agent/releases