Hermes Agent v2026.8.3 (Herald Release): what's new and how to use it in business

Nous Research shipped Hermes Agent v2026.8.3 – a version named the „Herald Release”. Hermes is an open-source (MIT-licensed) AI agent that we run our own company on, so we went through the changelog not for the list of 1,200 closed issues, but for one question: what of this actually changes for business. We picked five things that deliver concrete value – and for each we say how you can use it.

What's in this article?

1. What is Hermes Agent (in short)?
2. Voice that feels like a conversation
3. Agents talk to agents: the A2A v1.0 protocol
4. Hermes notifies your systems itself: signed webhooks
5. Research you can trust: citations and fact-checking
6. Less friction, longer autonomous runs
7. How we use Hermes ourselves?
8. Frequently asked questions (FAQ)

1. What is Hermes Agent (in short)?

Hermes Agent is an open-source AI agent from Nous Research – a tool that doesn't just chat but takes real actions: it runs commands, uses tools, integrates with company systems and works across many channels (terminal, desktop app, messengers). At our end it runs the company's daily work, which we wrote about in the piece on how we run our company with Hermes. Version v2026.8.3, from 3 August 2026, is one of the bigger releases – and the name „Herald” is no accident: in this version Hermes starts to speak, carry word to other agents and announce events to your systems.

2. Voice that feels like a conversation

Until now, voice mode in most tools worked like this: you speak, wait for the whole reply to generate, then listen to one long recording. In the Herald Release, Hermes speaks clause by clause as the answer streams, and you can interrupt it – cut in mid-sentence and it stops, listens, and knows it was interrupted. Add wake words detected locally, on-device (no audio leaves the machine until you say the phrase), plus voice on channels like WhatsApp, Slack or Feishu – send a voice note, get an answer.

How to use it in business: this is the foundation for a voice assistant and helpdesk that feels like talking to a person, not a rigid IVR menu. The customer calls, interrupts, asks follow-ups – and the agent responds naturally. It's exactly the direction we described around our AI Helpdesk and a voice layer for ticket handling.

3. Agents talk to agents: the A2A v1.0 protocol

A new, bundled plugin adds support for the Agent-to-Agent (A2A) protocol, version 1.0. In practice it means Hermes can discover other agents, talk to them and be driven by them – over a standard wire, regardless of the stack a given agent runs on.

How to use it in business: if you're building a multi-agent system from components by different vendors, A2A gives them a common language. Instead of hard-wired "everyone-to-everyone" integrations, you get one protocol over which a support agent, a sales agent and a research agent can talk. We described what such systems look like under the hood in the anatomy of a production multi-agent system – A2A is the missing standard connector between such blocks.

4. Hermes notifies your systems itself: signed webhooks

Previously, integrating with Hermes meant polling it or listening on a specific platform. Now Hermes pushes signed events itself (session activity, turn completions, tool events) to any HTTP endpoint you register – with an HMAC signature, so your receiver can verify it's really Hermes.

How to use it in business: it's a simple way to wire the agent into existing processes with no polling loop – a dashboard notification, a CI/CD entry, an action in an automation system, a chat alert. For us it's a natural extension of what we do in process automation and system integrations: the agent becomes an active part of the event flow, not a separate island.

5. Research you can trust: citations and fact-checking

The new grounded-citations skill makes Hermes produce research where every claim is backed by a verifiable source: quotes are matched against the actual page text (not made up), and links point to the specific evidence. A fact-checking mode turns the same machinery on any document – it tells you what checks out, what doesn't and what couldn't be verified.

How to use it in business: it's the difference between "sounds credible" and "has a documented source". For analyses, proposals or documentation you get material you can verify, not just a nicely worded paragraph. It's also a practical answer to the problem we described in the piece on why AI "lies" and how to fix hallucinations – forcing sources is one of the most effective tools. It lines up with what we wrote about AEO: in a world where AI cites sources, the verifiability of your content becomes currency.

6. Less friction, longer autonomous runs

A large part of the release is invisible but business-critical work on reliability:

• Tools that fix themselves. Truncated output spills to a file the agent can read back; an already-applied edit returns "ok" instead of an error; searches that match nothing suggest near-misses. Fewer wasted turns on tool friction.
• The iteration limit raised from 90 to 500. Long, autonomous tasks stop hitting an artificial wall and can finish the job.
• Correcting the agent on the fly. Instead of stopping and re-explaining, you drop a correction mid-run – the active turn changes course and the work so far isn't lost.
• Smarter approvals. Hermes can suggest an allowlist from your approval history, and a circuit breaker stops a runaway loop – less clicking "approve" without giving up control.

How to use it in business: these are exactly the things that decide whether an agent is fit for production. Longer tasks finished without supervision, fewer manual interventions and retained human control are the condition for trusting an agent with a real process – exactly how we approach AI agents as digital coworkers.

7. How we use Hermes ourselves?

We're not writing about this from the outside. Hermes runs our company's daily processes, and we extend it for our own needs – among other things we wrote it a plugin for our team chat that turns it into a virtual project manager. The Herald Release news plugs straight into that:

• voice feeds the direction we develop in the AI Helpdesk,
• A2A and webhooks are new ways to wire agents into the rest of your systems, which we handle in integrations and automation,
• citations raise the quality of the material the agent prepares.

If you want to put an agent like Hermes to work in your company – from the idea, through implementation, to integration with your systems – that's exactly what we do as part of AI implementations.

Frequently asked questions (FAQ)

The Herald Release shows where AI agents are heading: from a text assistant toward something that speaks, works with other agents and actively notifies your systems. If you want to use this in your company, take a look at our AI Helpdesk project, check out AI implementations or simply get in touch.