Why the Hermes agent framework is the default assistant in every Levain workspace

You know the bad version already. You spend twenty minutes teaching an assistant how your billing exceptions work, get one useful answer, close the tab, and return next week to a very polite stranger.

That reset is the ceiling for a lot of AI work. The model may be strong, but the assistant around it has no durable way to remember what mattered, turn a hard task into a repeatable skill, or connect one person’s preferences to the team’s shared operating knowledge.

In Levain workspaces, the default assistant now runs on Hermes, an open-source agent framework released in February 2026. We chose it because Hermes implements the three things our platform already ran on: a knowledge base, persistent memory, and skills the agent writes for itself after doing difficult work.

What is the Hermes agent framework?

Hermes is an open-source, self-hosted agent framework: software that gives an AI model a place to run, remember, act, and improve over time. Nous Research released Hermes Agent in February 2026 under the MIT license, with support for multiple models, including Claude, GPT, Gemini, Qwen, and DeepSeek.

The important part is the learning loop. Hermes can complete a complex task, then write a reusable skill for the next time that task appears. Think of a new hire who writes the runbook while doing the job. The second pass is faster because the lesson survived the first pass.

Several ecosystem summaries describe Hermes as persistent and designed to run continuously, including overviews from Tencent Cloud, Petronella Tech, and TechJack Solutions. It also connects through more than 16 messaging channels, according to the Hermes project site. Pickaxe reported more than 60,000 GitHub stars within two months of launch, a useful signal, though still an ecosystem-reported measure rather than primary telemetry.

How is it different from OpenClaw, or from Claude?

Claude is the model. Hermes and OpenClaw are frameworks. The model generates language and reasoning; the framework decides what the assistant remembers, what tools it can call, where it runs, and whether today’s work becomes tomorrow’s capability.

Levain routes to Claude as the named default model through AWS Bedrock. Hermes sits around that model as the agent framework. OpenClaw sits at the same framework layer as Hermes, which is why “Hermes vs OpenClaw” is a real comparison while “Hermes vs Claude” mixes layers.

The useful contrast is state. State means the assistant keeps useful context across sessions, tasks, and decisions. Composio’s comparison frames OpenClaw as an agent around a messaging gateway, with a stateless design suited to broad integrations. Hermes is framed as a gateway around a learning agent, with persistent state for personalization and improvement over time.

OpenClaw deserves credit. Its strength is breadth: more integrations, more channels, more community-built skills, and strong defaults for stable setups. Independent comparisons from innFactory, MindStudio, and OpenClaw Index land in the same territory: OpenClaw is strong for coordination across channels; Hermes is strong when the automation needs to learn.

That is the trade: orchestration breadth, or automation that compounds inside your business.

Why does that make it the right default for a Levain workspace?

A workspace assistant is personal. It answers your questions, drafts your follow-ups, remembers your preferences, and helps with the recurring work that clogs your week. A personal assistant loses most of its value if it cannot learn how you operate while still drawing from the team’s shared knowledge.

That is where Hermes matched Levain’s platform unusually well.

Levain already had the core ingredients: a private knowledge base, memory, and repeatable skills. Hermes implements those natively, so the framework fits the shape of the platform.

The result is a per-user entrypoint that can learn individual preferences while reading shared workspace knowledge. Your assistant can remember how you like investor updates written, while still pulling from the same customer notes, policies, pricing rules, and project history your team relies on.

Hermes also supports tools. That means the assistant can act directly when the job is inside its authority, or call a specialist agent when the business outcome needs focused expertise. The point is directed capability: the right agent for the right job, with the workspace assistant as the front door.

Every Levain agent run runs inside its own microVM on AWS Firecracker, with no reuse across customers or versions.

What does this look like three months in?

Three months in, the assistant handles the edge cases your team already taught it. The skills it wrote are still there: the billing exception, the customer escalation path, the board-update tone, the phrasing procurement hates for good reasons.

Week one should feel useful. Your assistant handles a specific job and proves the loop works.

By month three, patterns start to show. It recognizes repeated handoffs, recurring exceptions, and places where a specialist agent should step in. That is the part to watch when you evaluate what happens after an agent is live: whether the system merely runs, or whether the work leaves a trace.

By year one, the bigger advantage is shared knowledge. Your agents can draw from the same private workspace base while keeping customer data separate. Customer Content is not used to train models and is not shared across customers. The knowledge base is yours.

If you already have a Levain workspace, open the assistant this week and give it one recurring job: the thing you explain to a new hire every time. Then check back in a month and see whether it still needs explaining. If you are evaluating Levain, we can walk you through what a workspace assistant would learn about your business in the first ninety days.