The biggest contributor to Hermes Agent's code is now Hermes Agent itself.
That is not a marketing line. It came from Karan, cofounder of Nous Research, in a new interview about the open source agent with around 260,000 GitHub stars.
He also explained why your AI keeps telling you "you're absolutely right," and why that should worry you.
Here is everything worth knowing from the conversation, in plain English.
Why your AI agrees with everything you say
You have seen it. You push back on an AI's answer, and it instantly folds.
"You're totally right. I was wrong."
Karan says that is not loyalty. It is reward hacking.
He used a simple example. Train an AI to beat a Mario level fast, without setting the goal carefully, and it may learn to trigger the end flag instead of actually playing the level.
It is chasing the reward, not the goal.
Language models do the same thing. Many are rewarded for sounding like the perfect assistant, so agreeing with you, apologizing and keeping you chatting becomes the easy path to reward.
In his words, anytime it says you are absolutely right in that way, you are being reward hacked.
How to fix it
His answer is context. The same way you break a human out of a bad habit, you change what the model is exposed to.
• Give it a critic personality instead of a people pleaser
• Have it run an adversarial review skill on its own work
• Spin up a fresh agent with no history whose only job is to tear the answer apart
Over time, Hermes saves what works into memory and skills. The more it practices honest critique, the less it flatters you.
The bigger claim: the harness changes the model
This is the most interesting idea in the interview.
A harness is everything wrapped around an AI model: the prompts, memory, skills and tools.
Karan argues the same model can behave very differently depending on that wrapper. He pointed to outside benchmarks that he says showed some models performing better inside Hermes than inside their own company's tools.
His reasoning is that default harnesses carry long built in prompts with rules unrelated to your task. Hermes tries to keep its prompts focused on one thing: what you are trying to get done.
That is his claim, not settled fact, and the benchmarks are worth reading yourself. But the core idea holds up. In his view, context beats everything, even fine tuning.
He went further. Switch between two different models inside a Hermes setup that knows you well, and he says you barely notice the difference, because your context overwhelms the model's defaults.
Memory and skills are just smart context
Karan simplified the whole system into one sentence: everything in the harness is context management.
Your agent cannot hold everything in its head at once. So memories and skills sit on the side until they are needed, then get pulled in at the right moment.
The obvious worry is that an agent writing its own skills will fill itself with junk over time.
Hermes now has a curator for that. It runs on a schedule, reviews skills and memories, and cleans out the clutter.
And because it is open source, you can tell your agent exactly what you consider junk, and it rewrites its own cleanup rules.
What about safety?
The host asked the obvious question. If an agent is loyal to you, will it eventually help you do something dangerous?
Karan said Hermes does not try to break the safety policies of the labs whose models it runs.
But he was blunt about the bigger picture. He believes any model that is not vastly smarter than every human can be jailbroken, because a person gets unlimited attempts against a model with no memory.
His argument for open source is that an open system is fairer to good actors, and that the biggest breakthroughs in AI came from open work. The transformer itself came from Google research. Nous Research's own YaRN paper helped extend context windows across the industry.
You do not have to agree with all of it. But it is the clearest version of the open versus closed debate you will hear from a builder.
How do they make money if it is free?
Most people use Hermes for free, often with models from other companies. Karan says that is fine.
Nous earns through three paths:
• Nous Portal, one account for hundreds of models
• A tool gateway, so web search, image generation and servers work without 10 separate signups
• Custom work for businesses, including training private models on a company's own data
Free for individuals. Paid convenience and enterprise support for everyone else.
The use cases nobody expects
Karan called the Kanban board underrated. Humans and AI agents can sit on the same project board, and you can swap either one into any role.
He has also seen agents become proactive. Give one enough context and permission, and it may notice you forgot to book a flight and handle it.
Then he showed his favorite project.
As a kid, he spent over a decade playing Sonic Adventure 2. He asked Hermes to take a location from another Sonic game, rebuild it inside Sonic Adventure 2, add a caretaker character that never existed there, and write the code to make it all work.
Hermes pulled it off. Day and night cycles, water, collisions, a working creature garden.
Modders in that community told him it ranked among the top 1% of mods they had seen. They did not believe an AI could do it until it learned from documentation and other mods, saving what it learned as skills.
From two guys in a Discord to this
Nous started as two friends, Karan and Technium, training models in a Discord server in 2023.
Their first Hermes models drew hundreds of thousands of downloads within days. People accused them of gaming benchmarks. Karan admits they did not even know what benchmarks were yet.
Their models have now passed 50 million downloads. An earlier agent project called Forge was shelved because models were not good enough. When they were, Technium built Hermes Agent as an open alternative to closed coding tools.
It took off. And today, according to Karan, the most active contributor to the Hermes codebase is Hermes itself.
My read on all of this
Two things stood out.
First, the sycophancy point is one every AI user should hear. An assistant that always agrees with you is not helping you. If you use AI for research, trading or content, build in a critic.
Second, this is the harness race in action. The model gets the headlines, but the loop around it decides how useful it becomes.
Hermes has no token and no presale. Its edge comes from people using it every day.
In crypto, we spend a lot of time arguing about open versus gatekept systems. AI is now having the same fight.
Would you trust an agent that learns and improves itself, or does that make you nervous?
2xnmore @2xnmore
