Overview of Hermes Agent: Agents that grow with you
This episode of the Practical AI Podcast features Jeffrey Cannell, co-founder and CTO of Nous Research, discussing how the company evolved from an open-source research community into a product-building organization centered on Hermes Agent—an agentic system designed to improve as it is used. The conversation covers the state of open source AI, the rise of agent harnesses, why workflows are shifting from model-centric to outcome-centric, and how agent memory and skill accumulation can create compounding value for both individuals and organizations.
Key Takeaways
- Nous Research grew from a loose open-source collaboration into a formal company focused on keeping AI accessible and open.
- Their early work included:
- the Hermes line of aligned open models,
- research into distributed training over the internet,
- and efficiency-focused academic work aimed at lowering AI costs dramatically.
- Hermes Agent began as an internal tool for model research and later became open source after the team realized it had strong product-market fit.
- The core idea behind Hermes Agent is that the agent gets better the more you use it by accumulating:
- skills
- memory
- and contextual knowledge about user preferences and recurring tasks.
- The discussion frames AI as a human capability multiplier, especially for small teams and individuals with strong judgment and vision.
- The hosts and guest repeatedly stress that outcomes matter more than model choice alone in agentic systems.
The Evolution of Nous Research
From Discord community to company
Jeffrey described Nous Research as starting with a community of open-source AI enthusiasts who wanted to ensure powerful AI wasn’t locked behind a few closed companies. What began as informal collaboration eventually became a company so that people could work on this mission full-time.
Research as the first “distribution strategy”
Before Hermes Agent, Nous focused heavily on research that could create major efficiency gains—things they describe as “1000X hacks” that could bring expensive AI capabilities into reach for smaller players.
Hermes models before Hermes Agent
Their original Hermes model series became well known for producing models that felt more aligned to users and less preachy or moralizing. That success helped establish Nous as a credible open-source AI research group.
Open Source AI in the Current Landscape
The role of Meta/Llama
Jeffrey notes that Llama was crucial in bootstrapping the open-source AI ecosystem, especially for researchers and builders who got started by downloading and experimenting with those models.
Why open source became geopolitically important
The conversation highlights how open source AI moved from being a technical and economic issue to a geopolitical one, especially after strong open-source models began emerging from China. That shifted the center of gravity and made the competition more strategic.
NVIDIA’s involvement
One of the most important recent developments, according to Jeffrey, is NVIDIA/Jensen Huang’s commitment to support western open-source model development. He sees this as potentially transformative because it aligns directly with NVIDIA’s own business interests: more open models ultimately means more demand for NVIDIA hardware.
What Hermes Agent Is and Why It Matters
The “brain and body” analogy
Jeffrey’s central framework is simple:
- The model is the brain
- The harness is the body
The model provides intelligence and reasoning, but the harness is what allows the AI to act in the world—touching systems, using tools, persisting state, and producing outcomes.
Why the harness matters more now
As agentic systems become more common, the harness becomes increasingly important because:
- it connects the model to tools and workflows,
- it handles memory and state,
- and it turns abstract reasoning into real-world action.
Value shifts toward outcomes
Jeffrey argues that the market will increasingly pay for completed outcomes, not just raw model access. In his view, the future is less about “what model are you using?” and more about “what job got done?”
How Hermes Agent Is Designed
Built to get better with use
The defining principle of Hermes Agent is that it should improve through usage rather than requiring humans to predefine every workflow.
Two core systems
1. Skill system
- The agent observes repeated or successful behaviors.
- It creates reusable “skills” automatically.
- Future tasks can reuse those skills without manual reimplementation.
A memorable example from the episode:
- Jeffrey used Hermes Agent to book a restaurant.
- The system had to work through bot protections, discover the booking backend, and then complete the reservation.
- Afterward, it created a skill for that kind of task, so the next reservation was much faster.
2. Memory system
- Hermes Agent builds layered memory about:
- how to treat the user,
- prior conversations,
- and recurring preferences.
- It uses this memory selectively, based on its own judgment.
Opinionated design choices
The team intentionally keeps the hard-coded feature set small:
- code execution
- web browsing
- basic tool access
Everything else is intended to emerge through the model’s own self-reflection and learned behavior.
Best Use Cases and Market Fit
Who it’s for
Hermes Agent is positioned first for:
- power users
- people who want to push AI to its full capability
- developers and advanced users who value open-source native workflows
Deployment options
It is designed to work:
- locally with open models
- or as a hosted service
That flexibility makes it appealing to users at different points on the openness/compliance spectrum.
Growing into workplace and team settings
The team is now exploring:
- multi-agent workflows
- collaborative use cases
- enterprise applications
Jeffrey shared an example of an internal agent connected to backend infrastructure via MCP that began as a debugging tool and gradually accumulated enough knowledge to support:
- customer support
- root cause analysis
- infrastructure inspection
- account-level troubleshooting
This illustrates the compounding advantage of letting agents learn from real organizational work.
Human Work, Automation, and Retraining Ourselves
Don’t automate the human process one-to-one
A major theme in the discussion is that the best automation often does not mirror the human workflow directly. Agentic systems should be designed around the actual operational problem, not the way a human would normally solve it.
Think of agents as:
- infinite patience
- but limited creativity
This makes them ideal for tasks that are:
- repetitive
- log-heavy
- time-consuming
- evaluation-driven
- not heavily dependent on human taste or originality
Outcome-first prompting
Jeffrey recommends:
- describing the desired outcome
- defining the success criteria
- and letting the agent figure out the execution path
This is a major shift from micromanaging each step.
Aesthetic judgment is still human
The episode also emphasizes that current models often struggle with:
- taste
- aesthetic discernment
- and the subtle “this feels right” judgment humans bring to work
Models can be brilliant at computation and reasoning, but still produce results that feel “sloppy” or unnatural to people unless explicitly guided.
Broader Implications for the Future
AI as a lever for individual impact
Jeffrey is strongly optimistic that a small number of motivated people can now have outsized influence because AI amplifies individual capability.
The human side of AI adoption
He also expresses concern that people may:
- lose critical thinking
- over-delegate judgment to agents
- or grow up in a world where “the computer knows” is the default answer
This is especially important for younger generations who may never have known a pre-agentic world.
Nous Research’s mission
The company frames its work as human-centric AI:
- better yesterday than today
- better today than tomorrow
That includes not just building better tools, but helping people use them in ways that preserve agency and human growth.
Practical Advice from the Episode
- Focus on outcomes, not steps.
- Be explicit about evaluation criteria.
- Use agents where patience matters more than creativity.
- Let systems accumulate knowledge over time.
- Treat open source as a strategic and community-building force, not just a distribution tactic.
- Expect agentic systems to change how you work, not just how you code.
Closing Thoughts
This episode positions Hermes Agent as more than just another agent framework—it’s a bet on systems that learn from use, retain useful memory, and become more valuable over time. The larger message is that the AI industry is shifting from models as isolated intelligence engines to agentic systems that turn intelligence into action, and that shift opens major opportunities for both open-source builders and individual practitioners.