Most coding agents forget everything the moment you close the terminal. Hermes Agent, the open-source framework from Nous Research, tries something different: it creates skills from experience, improves them during use, nudges itself to persist knowledge, searches its own past conversations, and builds a deepening model of who you are across sessions.
That claim is either the most important thing happening in open-source agents right now, or a very sophisticated CLAUDE.md wrapper. After reading the primary sources and tracking the release changelog, the answer is somewhere specific - and worth understanding precisely before you commit to it.
What the Hermes learning loop actually does
Hermes Agent is an autonomous AI agent with a built-in learning loop. It creates skills from experience, improves them during use, searches its own past conversations for context, and builds a deepening model of who you are across sessions. Those are the marketing words. Here is the mechanism.
The agent can create, update, and delete its own skills via the skill_manage tool. This is its procedural memory - when it figures out a non-trivial workflow, it saves the approach as a skill for future reuse.
Skills and memory work together in the self-improvement loop: memory stores small durable facts that should always be in context, while skills store longer procedures that load only when relevant.
The trigger threshold matters here: after complex tasks involving 5 or more tool calls, the agent creates reusable skill documents, and those skills are patched during use when they are outdated, incomplete, or wrong.
By default the agent writes skills freely, including from the background self-improvement review that runs after a turn. If you would rather approve every skill write first - useful for small models that misjudge what they learned, secure environments, or just wanting eyes on the self-improvement loop - there is a write-approval gate you can enable.
That approval gate is the honest answer to whether this is safe to run in a production environment. It is not on by default. If you enable it, skill updates are staged for human review before committing. The self-improvement loop without that gate is genuinely autonomous - which is either exciting or alarming depending on your risk appetite.
The evolutionary self-improvement tooling operates entirely via API calls - mutating text, evaluating results, and selecting the best variants.
A full optimization run costs roughly $2-10.
What v0.18 actually changed - and why it matters
Hermes Agent v0.18.0 landed on July 1, 2026 - twelve days after v0.17.0. The team called it "The Judgment Release," and the name carries double weight: first, a judgment call on the backlog (the team resolved every open P0 and P1 issue and PR in the entire repository before cutting the release), and second, better judgment in the agent itself - how well it thinks, and how it knows its work is actually done.
The release window covers approximately 1,720 commits, 998 merged PRs, 2,215 files changed, and 949 issues closed, with contributions from 370+ community members.
Two specific additions from v0.18 stand out.
Self-verifying goals. The agent now records verification evidence for coding work and can decide it is finished by actually running your project's checks, not by asserting success.
The /goal command gained completion contracts: you state what "done" looks like, and the standing-goal loop judges completion against that evidence instead of stopping when the model feels like it.
This is a direct attack on one of the most common failure modes in long-running agents - the confident hallucination of completion.
Mixture-of-Agents as a first-class model. MoA became a first-class citizen - named ensembles of models you can pick like any other model, with every reference model's reasoning shown to you and the aggregator's answer streamed live.
You read what GPT-5 thought, what Claude thought, what Grok thought, then watch the aggregator synthesize them into an answer that streams live. The inspectability here is genuinely new - most MoA implementations give you the merged answer and hide the individual reasoning.
Running Hermes in Slack or Microsoft Teams
The gateway supports Telegram, Discord, Slack, WhatsApp, Signal, DingTalk, SMS, Mattermost, Matrix, Webhook, Email, Home Assistant, Microsoft Teams, Google Chat, LINE, and SimpleX Chat, all running from one gateway process. The breadth is real. The setup complexity varies significantly by platform.
For Teams specifically: Hermes v0.14.0 (released May 16, 2026) added native Microsoft Teams support as its 20th messaging platform. The integration includes Microsoft Graph authentication, webhook-based message ingestion, the standard Hermes agent pipeline, and outbound delivery via Graph API. Setup requires an Azure AD app registration, a public HTTPS endpoint, and explicit user allowlisting.
With Azure experience, expect 1 to 2 hours for basic bot setup and testing. Without Azure experience, plan for 3 to 5 hours.
One sharp edge:
the most common cause of silent failures is the TEAMS_ALLOWED_USERS list. Hermes drops messages from users whose AAD object ID is not in the allowlist, with no visible error.
That will waste an afternoon if you do not know it going in.
The Slack setup follows the same gateway pattern - one process, webhook ingestion, same agent loop. A teammate like Beagle takes a different approach: it runs natively inside Slack with no Azure registration, no public endpoint, and no tunnel required, which is the right tradeoff for teams who want AI in their channels without owning the infrastructure.
Beagle in action#eng-ops, 10:40amThe honest hype check
Hermes Agent is the most sophisticated open-source agent framework available right now. It has a closed learning loop with skill creation and curation, persistent memory, 20+ messaging platforms, six terminal backends, cron scheduling, MCP integration, and isolated profiles. That is a real platform, not a project.
What it is not: a production-ready managed service. The learning loop is powerful, but by default the agent writes skills freely - including from the background self-improvement review. If you want a human reviewing every skill write, you must configure the write-approval gate explicitly. Teams running Hermes at scale need to treat that gate like a security control, not an optional setting.
The naming is also genuinely confusing. Hermes 4 is the large language model. Hermes Agent is the self-improving agent framework that runs on top of a model. Hermes Desktop is the graphical app for that framework. Three different things under one brand, and the distinction matters when you are evaluating capability versus reading release notes.
The star count - 215k+
- is meaningful but also a product of the OpenClaw migration path: if you are coming from OpenClaw, Hermes can automatically import your settings, memories, skills, and API keys, and the setup wizard detects an existing OpenClaw installation and offers to migrate before configuration begins. Some share of that star count is people following the migration trail, not independent discovery.
The v0.19 release (July 20, 2026) gave one more concrete signal about where the project is heading: first-turn time-to-first-token dropped approximately 80% on every platform, reasoning streams live by default, the desktop app got a speed overhaul with 14× faster streaming markdown, and the TUI renders markdown incrementally. That is a project attacking production-readiness problems, not just feature shipping.
Hermes Agent open source: common questions
What is Hermes Agent and how does it differ from Hermes 4?
Hermes Agent is an open-source autonomous agent framework by Nous Research with a built-in self-improvement loop. Hermes 4 is Nous Research's language model. The two are separate products that share a brand: Hermes Agent runs on top of any LLM provider - Claude, GPT-5, local models - and Hermes 4 is one option you can point it at.
Does the Hermes Agent self-improvement loop actually work without retraining?
Yes, and the mechanism is worth understanding: skills are stored as plain markdown files. After a task requiring five or more tool calls, the agent writes a skill document describing the workflow. On future similar tasks, it loads that skill and patches it if it finds a better approach. No GPU, no fine-tuning - it is prompt engineering that persists. The ceiling is the base model's reasoning quality.
How hard is it to connect Hermes Agent to Microsoft Teams or Slack?
Teams is the harder integration. You need an Azure AD app registration, a public HTTPS endpoint (or a dev tunnel), and explicit user allowlisting by AAD object ID. With Azure experience, expect one to two hours. Slack follows the same gateway pattern and is somewhat simpler. Both require self-hosting Hermes and managing the infrastructure yourself.
What did Hermes Agent v0.18.0 actually add?
Three things worth knowing: self-verifying goals (the agent runs your project's actual checks before declaring a task done, not just asserting completion), first-class Mixture-of-Agents with inspectable per-model reasoning, and the /learn command for explicit skill distillation. The release also cleared 100% of open P0 and P1 bugs - roughly 692 highest-priority issues in twelve days.
Is Hermes Agent production-ready for a team?
Capable, yes. Managed and low-maintenance, not yet. The platform requires real infrastructure ownership - self-hosted gateway, Azure registration for Teams, manual security configuration. The self-improvement loop needs the write-approval gate enabled for any production context. It is the right choice for teams with a developer willing to own the deployment. For teams without that, a managed integration is a more realistic starting point.