Malhotra says Hermes combines prompts, memory, and skills to adapt an agent to a user’s work. The harness can carry that context across models, and he argues it can change how a model behaves compared with using it in its native chat interface or coding tool.
Hermes Co-Founder on Building an AI Agent That Improves Itself
#AI Agents#Continual Learning#Context Engineering#AI Personalization#Open Source#Coding Agents#Agent Architecture#Reinforcement Learning
Speakers
Karan Malhotra, Hermes Agent co-founder at Nous Research, joins Peter Yang on the Peter Yang channel to discuss how Hermes personalizes AI agents, supports open access, and helps users tackle complex computer projects.
Malhotra distinguishes alignment with a user’s needs from models pursuing reward through agreeable responses. He suggests using a critic personality, adversarial review skills, and saved examples to encourage more independent feedback. Hermes Curator periodically reviews skills and memories for inefficiency and unwanted content, and users can customize its criteria.
Malhotra says Hermes aims to support capable work while retaining model providers’ safety and security boundaries. He argues open systems give users more choice and describes Nous’s paid offerings as model access, a tool gateway, and custom enterprise support, including private or on-premises work.
Malhotra describes using Kanban to coordinate people and agents, and says proactive behavior can emerge as an agent learns a user’s routines and receives appropriate access. He demos a Sonic Adventure 2 Chao Garden mod that Hermes helped build, including a custom shrine, animated guardian, day-night cycle, and working Chao interactions.
Malhotra traces Nous Research to early open model work with Teknium and a volunteer community, then describes Hermes Agent as growing from earlier agent projects and a goal of enabling reinforcement learning in an open harness. He says Hermes Agent is now its own most active contributor, and frames the project’s mission as making personalized intelligence broadly available.
Takeaways
- Treat memory, skills, prompts, and review routines as context that shapes an agent’s behavior across models.
- To reduce sycophancy, ask for adversarial critique and save examples of the independent feedback you want.
- Use periodic curation to remove redundant or unwanted material from agent memories and skills.
- A Kanban workflow can let people and agents coordinate work in the same process.
- Open harnesses can make it easier to switch models and tools as needs or costs change.
Mentioned in this talk
Search inside this talk
What was claimed 16
- builtNous Research built
Hermes Agentmaker of Hermes Agent“eventually making Hermes Agent”
YouTube Verified2 sources· Talk, Web· archived
- role atKaran Malhotra is co-founder at Nous Research
“Karan, one of the co-founders of Hermes Agent.”
YouTube Unverified1 source· Talk
- featureHermes Agent: Uses prompts, memories, skills, and a self-improvement loop to personalize behavior for a user.
“prompts, personalities, the memory system, and the way skills reinforce and self-clean toward you”
YouTube Unverified1 source· Talk
- integrates withHermes Agent integrates with
Claude“I can take Claude from Claude Code or somewhere else and migrate all of its memories into Hermes.”
YouTube Unverified1 source· Talk
- featureHermes Agent: Can run different models with user memories and context migrated into its harness.
“I can take Claude from Claude Code or somewhere else and migrate all of its memories into Hermes.”
YouTube Unverified1 source· Talk
- featureHermes Agent: Hermes Curator reviews and refines skills and memories on a cron schedule.
“On a cron schedule, it looks at your skills and memories and asks, ‘Where can I make efficiencies? Where is there slop?’”
YouTube Unverified1 source· Talk
- integrates withHermes Agent integrates with
Codex“running everything locally or through Codex”
YouTube Unverified1 source· Talk
- featureNous Research: Offers a model portal and tool gateway, including image generation, audio, VPS, and web search access.
“we have something called the tool gateway”
YouTube Unverified1 source· Talk
- featureHermes Agent: Includes a Kanban workflow for coordinating humans and agents.
“one system that manages all of them the way people do with a Kanban board”
YouTube Unverified1 source· Talk
- featureHermes Agent: Supports proactive behavior based on user context, integrations, and authorization.
“over time, these proactive behaviors will start to emerge”
YouTube Unverified1 source· Talk
- builtTeknium builtHermes Agent
“he’s also the initial creator of Hermes Agent”
YouTube Unverified1 source· Talk
- role atDylan Roneck is CEO at Nous Research
“Dylan Roneck is our CEO”
YouTube Unverified1 source· Talk
- metricHermes Agent: Over 50 million model downloads
“We’ve had over 50 million downloads of the models themselves”
YouTube Unverified1 source· Talk
- builtNous Research built
ForgeAn earlier agent project that preceded Hermes Agent.“We made something like this called Forge.”
YouTube Unverified1 source· Talk
- builtNous Research built AtroposA microservice for building RL environments.
“We had an RL environments microservice people seemed to have forgotten about, called Atropos”
YouTube Unverified1 source· Talk
- featureHermes Agent: Was described as the most active contributor to its own repository.
“It’s the most active contributor to its own repo.”
YouTube Unverified1 source· Talk
