Hermes Agent · by Nous Research
The AgentThat GrowsWith You
Hermes is an open-source, self-hosted agent that coordinates frontier models to do real, multi-step work—and, unusually, it learns. It remembers, writes its own skills, and gets more capable the longer it runs. That compounding power is exactly what needs a guide.
What Makes It Different
Most agents forget. This one compounds.
A one-shot agent starts from zero every time. Hermes runs a learning loop—so the work you do together this week makes next week’s work faster.
Do the work
You give it a goal; it coordinates a model—often Claude—to execute the multi-step task.
Remember
It keeps memory across sessions, so context and outcomes carry forward instead of resetting.
Build a skill
When a task works, it writes a reusable routine from the experience—a new capability it keeps.
Improve
Next time the job comes up, it’s faster and sharper—because it’s done it before.
↻ and it repeats—getting sharper every cycle
The Power
A long-running agent on your own terms
Persistent memory
Hermes accumulates what it learns across sessions, so it doesn’t start cold every time you call on it.
Writes its own skills
It turns successful work into reusable routines—building a library of capabilities from real experience.
Model-agnostic
Point it at the model you trust—it commonly coordinates Claude via Claude Code—and swap as you like.
Runs on a schedule
As a long-running daemon it can fire scheduled and recurring jobs, not just answer in the moment.
Real tools
It executes tasks, searches the web, writes and runs code, manages files, and talks to external services.
Open-source & self-hosted
MIT-licensed and runs on your own infrastructure—your data, your models, your control.
Where Thinking Backward Comes In
An agent that learns
learns whatever you teach it
A self-improving agent is only as good as what it picks up. Left alone, it can just as easily compound bad habits, sloppy skills, and access it never should have had.
We coach you to stand Hermes up, shape what it learns, curate its skills, and set the guardrails—so it compounds in your favor. You own the agent and the judgment behind it. Capability that stays with you.
Hands-on, with you
You leave able to run, shape, and extend the agent yourself.
Curate what it learns
Good skills in, bad habits out—so the compounding works for you.
Guardrails first
Scoped access and review, because a learning agent shouldn’t learn the wrong reach.
What We Build Together
An agent that gets better on purpose
We start from the work you want compounding in your favor—and work backward to a Hermes setup that learns the right things.
A self-hosted Hermes setup
Installed on your infrastructure with the right model and credentials—stable, private, and ready to run long jobs.
A curated skill library
We seed and shape the routines it learns so it gets reliably good at the work that matters to you, not random tasks.
Memory & guardrail policy
What it remembers, what it can touch, and where a human signs off—so a self-improving agent stays trustworthy.
A compounding workflow
The recurring jobs—research, monitoring, ops—wired up so Hermes measurably saves you more time each month.