The Herald Release shipped on August 3, 2026 with ~3,650 commits, ~1,400 merged PRs, and ~1,200 issues closed since v0.19.0 on July 20. That is two weeks of work. Most agent frameworks don't move that fast in a year. The question worth asking before you set it up in your team's Slack is not whether Hermes Agent is active - it obviously is - but whether the thing it actually ships in v0.20 is the right fit for how your team works.
What actually shipped in Hermes Agent v0.20
The Herald Release turns the open-source agent from a text-based assistant into something that speaks, coordinates with other agents over a standard protocol, pushes signed events to external systems, and backs its research with verifiable citations.
The four headliners, in order of likely impact:
- Real-time conversational voice. Hermes now speaks clause-by-clause as its reply streams, and you can interrupt it mid-sentence by talking.
First token comes in around 0.9 seconds, down from 4.3 seconds in prior versions. This matters less for Slack-based workflows and more for desktop and CLI users, but voice replies now also work on messaging platforms.
- A2A v1.0. A2A is a new plugin that lets Hermes speak the agent-to-agent protocol - discovering, talking to, and driving other agents.
In plain terms, Hermes can become the front door to your whole agent stack, or a worker inside someone else's. This is the sleeper feature of the release - not useful on day one, but worth watching as the A2A ecosystem grows.
Grounded citations. A new research skill matches claims against actual page text and can fact-check any document you hand it. Useful for anyone using Hermes for synthesis tasks rather than code execution.
Signed outbound webhooks. Hermes now pushes lifecycle events to any HTTP endpoint you register, instead of requiring you to poll it. This is the feature that quietly makes Hermes a real automation node, not just a chat interface.
How Hermes Agent Slack sessions actually work
The recommended platform for shared team use is Slack. Slack sessions are keyed by thread, not by user - multiple users in the same thread share one conversation, which is exactly the behavior you want for a team agent.
That is the design, and it is genuinely useful. But it carries a consequence most guides skip: the agent's memory and context are shared too. If your teammate tells Hermes something in the thread - a project constraint, a correction, a preference - it stays in that session. There is no per-user memory isolation within a thread. Memory stores facts: things the agent knows about you, your projects, and preferences, retrieved automatically based on relevance. Skills store procedures - step-by-step instructions for how to do things - recalled when the agent encounters a similar task. Both persist across sessions. In a shared thread, that persistence is collective, not individual.
The practical consequence: Hermes works well as a team-visible agent in a dedicated channel - a #research-agent or #ops-agent channel where everyone understands the shared context. It works less well if you're treating it as a personal assistant that happens to be visible to colleagues.
Beagle in action#product, 10:22amThe model-agnostic angle is the real differentiator
Hermes Agent is a fully open-source agent framework that lets you plug in any inference provider you want. That is the sentence that separates it from Claude Code, Codex, or most commercial agent products.
The Nous Portal acts as the company's "unified subscription gateway" - a single login provides access to more than 300 models from multiple providers, including Claude, GPT, Gemini, and DeepSeek models, plus a "Tool Gateway" that routes web search, image generation, text-to-speech, browser automation, and terminal sandbox services through the same subscription.
As of July 2026, hosted paid tiers ranged from $20 to $200 per month.
Here is what that means in cost terms. DeepSeek V4 Flash charges $0.28 per million output tokens. That is a 643x gap between that model and GPT-5.5 Pro, which you can call with nearly identical request bodies. An agent routing its lighter tasks - lookups, summaries, ticket drafts - to a cheap flash-tier model and its harder reasoning to a frontier model can cut a meaningful fraction of its inference bill without changing the agent software at all.
| Task type | Suggested model | Approx. output cost/M |
|---|---|---|
| Lookup, summarize, route | DeepSeek V4 Flash | $0.28 |
| Code drafts, PR review | GPT-5.6 Terra | $12 |
| Complex agentic reasoning | GPT-5.6 Sol / Claude Opus 4.8 | $25-30 |
| Batch reports, evals | Any model via Batch API | ~50% discount |
The sticker price almost never tells you what you'll actually pay: caching, batching, and long-context surcharges swing real costs 50-90% in either direction. Hermes's provider-agnostic model picker - now fuzzy-searchable across every interface in v0.20 - is where that optimization actually happens.
What Hermes Agent v0.20 does not do for teams
Be honest about two gaps before you deploy.
First, the Hermes model and the Hermes agent are different products. Hermes Agent is Nous Research's open-source AI agent, released under the MIT License and developed in public on GitHub. The Hermes 4 model family - the actual LLM weights - is a separate thing. Hermes 4 70B is an open-weight model from Nous Research with no published intelligence, coding, or agentic benchmark scores, making its real-world capability difficult to verify against current models. Most teams deploying Hermes Agent are pointing it at a different model entirely - Claude, GPT, DeepSeek. The naming confusion is real.
Second, the architecture is personal-agent-first. The agent is built around a persistent learning loop: it creates reusable skills from its own experience, searches its past conversations, and maintains memory of the user across sessions. "The user" in a Slack deployment is effectively the thread. That works for dedicated use cases. It becomes noisy when the thread is a general team channel where topics jump around. Create a separate Hermes channel per use case, not one catch-all bot.
A teammate like Beagle, built specifically for Slack and Teams, handles the approval layer and message routing that Hermes leaves to you to configure. The two aren't in direct competition - Hermes is an agent runtime you operate; Beagle is a layer that lives natively in your existing workspace. Worth knowing which problem you actually have before you pick one.
Hermes Agent Slack: common questions
What is Hermes Agent v0.20 and what's new?
Nous Research shipped Hermes Agent v0.20.0 on August 3, 2026 - described as the biggest single release in the project's history. The Herald Release turns the agent into something that speaks, coordinates with other agents over a standard protocol, and backs its research with verifiable citations. Key additions: real-time voice with barge-in, A2A v1.0 protocol support, grounded citations, and signed outbound webhooks.
How does Hermes Agent connect to Slack?
Hermes runs a single background gateway process that connects to configured platforms including Slack, Telegram, Discord, and WhatsApp, handling sessions, cron jobs, and voice messages.
You create a Slack bot app, supply the credentials, and run hermes gateway run.
Slack sessions are keyed by thread - multiple users in the same thread share one conversation context.
Is Hermes Agent free to run?
Hermes Agent remains MIT-licensed open source, and v0.20.0 is available immediately via the shell installer or pip install hermes-agent. Hosted tiers exist for those who prefer managed infrastructure, but the core agent is free to run on your own hardware.
You pay for inference tokens through whichever provider you configure.
Can Hermes Agent work with any LLM?
Yes. Hermes works with 20+ inference providers. You can switch from DeepSeek to Claude to a local model mid-workflow - one config change, no architecture change. The Nous Portal bundles access to 300+ models under a single subscription for teams that don't want to manage API keys per provider.
What is the A2A v1.0 feature in Hermes v0.20?
A2A allows two separate Hermes Agents to communicate directly, exchange tasks, and share context - so one agent scrapes raw data, another analyzes it, and a final summary is generated automatically. It implements the open Agent-to-Agent protocol, meaning Hermes can also interoperate with A2A-compatible agents from other frameworks. The full multi-agent orchestration story is still early, but the protocol foundation is now in place.