Hermes Agent: Agents that grow with you

Noose Research began roughly two years ago out of concern that AI capability would consolidate behind a small number of closed companies, a worry that felt urgent when GPT-3.5 and ChatGPT launched and the capitalization curve was already steepening. Early work focused on publishing research targeting thousand-times efficiency improvements on expensive AI pipeline components and releasing fine-tuned, non-moralizing models under the Hermes name. Kanell noted that top AI research engineers are now compensated like NFL quarterbacks and are as strategically critical as GPU access, while acknowledging that Noose's funding is a rounding error compared to the catering budget of larger AI organizations.

Hermes Agent started as an internal tool built roughly six months before recording to support Noose's model research team, with the earliest commits focused entirely on model training use cases. The original developer, known as Technium, had little to no prior coding ability and built it using AI tools, which the speakers cited as direct evidence that AI acts as a human capability multiplier enabling non-developers to ship production-grade applications. After Noose defaulted to open sourcing it, the project became the number one open source repository on GitHub and was reportedly written 99.99 percent by Hermes Agent itself.

Kanell described agentic architecture using a brain-body analogy: the model is the brain capable only of outputting tokens, and the harness is the body that gives the model persistence, causality, and forward-moving time. He argued that a better harness with a worse model can outperform a better model with a worse harness. Hermes Agent was intentionally designed with very few hard-coded features, giving the model minimum ability to touch the world, limited to running code jobs and browsing the web, and otherwise getting out of the model's way.

The memory and skill systems in Hermes Agent are emergent properties encouraged through prompts rather than hard-coded logic. A self-reflection mechanism prompts the model to identify frequently used patterns or important achievements and store them using the model's own judgment about what is worth remembering. When tasked with booking a Las Vegas restaurant reservation, the agent encountered bot-capture systems, eventually found the API backend, completed the booking after roughly thirty to forty-five minutes of real runtime, and automatically created a skill labeled Las Vegas restaurant booking for future use. Internally, Noose deployed a Hermes agent connected via MCP to all back-end infrastructure in a read-only capacity, and over one month of use by engineers it accumulated a large repertoire of company-specific skills without explicit programming, with learnings from one engineer becoming available to the entire organization including non-technical support staff.

Because the architecture relies on model judgment rather than hardcoded logic, the speakers predicted Hermes Agent will automatically improve when more capable models such as a future GPT or Claude Opus release, without requiring changes to the harness itself. The primary target market is power users who want to use AI to its full capability, and local model support is a first-class citizen alongside hosted deployment.

On the broader open source landscape, Meta's decision to open source Llama was driven largely by Zuckerberg's personal judgment rather than formal business strategy, and Llama bootstrapped the AI researcher ecosystem. Llama 4 was estimated to have cost between two hundred million and four hundred million dollars to train and was considered a failure due to architectural mistakes, prompting Meta to question its open source motivations. DeepSeek releasing a near-state-of-the-art open source model introduced a geopolitical dimension, and Chinese companies adopted open source releases as a growth hack that shifted the origin of such releases from predominantly Western to almost exclusively Chinese. At GTC, NVIDIA committed approximately twenty billion dollars over several years toward training Western open source models, which the speakers described as a potential game changer because NVIDIA's business interest is uniquely aligned with open source since all AI workloads ultimately run on its chips.

Agents are best understood as entities with infinite patience but very little creativity, suited for tasks any non-genius human could theoretically perform but that no human would realistically want to do. The wrong framing is mapping agents to human job titles such as CEO agent or CFO agent; the more productive approach asks what a task actually involves rather than what title it carries. Unstated assumptions about output quality are a persistent source of unsatisfactory results, and people who can articulate their internal suppositions succinctly provide disproportionate value because AI must be treated like an entity that never grew up on Earth and requires explicit communication of context that would otherwise go without saying.

Noose frames its positioning as human-centric AI, motivated by concern that AI agents could accelerate a societal trend toward disengaged thinking already visible in doom-scrolling behavior. Kanell noted he recently had a baby and is personally motivated by what the AI paradigm will mean for a generation that grows up with it as the default, warning that it takes only one generation for a new paradigm to become what people consider normal, and that such a generation may never develop independent critical thinking or recognize human innovation as valuable.