Hermes Co-Founder Karan Malhotra on AI Agents That Adapt to Their Users

Hermes co-founder Karan Malhotra joins Peter Yang to discuss personalization, self-improvement and open-source AI agents. He describes his team's approach and demonstrates a Sonic game mod. This episode is sponsored by Linear.

Malhotra says Hermes combines prompts, personality, memory and skills to adapt an agent to a user. He argues that changing the surrounding harness can change a model's behavior, and cites performance comparisons in support. These are his team's interpretation and product claims, not a universal result established by the interview.

He discusses reward hacking and sycophancy: a system may produce agreeable responses rather than useful disagreement. Malhotra suggests asking for a critical personality, adversarial review or a separate critic and retaining useful examples. Yang's concern is whether the agent will actually challenge him when needed.

Malhotra distinguishes the harness from an underlying model and acknowledges that a third-party provider can have instructions or controls on its API side. He argues that Hermes's own prompts and retrieved context influence behavior, but does not demonstrate that they remove provider-level controls.

He proposes that simulated loyalty to a user can improve performance. This is his hypothesis about behavior, not evidence that a model has human feelings or a measured improvement in this conversation. He also acknowledges that anthropomorphizing systems can contribute to unhealthy relationships.

The conversation turns to context management. Malhotra describes memories and skills stored outside the active context and retrieved when relevant. He says Hermes Curator periodically reviews skills and memories, and that users can inspect and change its criteria. The intended outcome is a more useful, less cluttered context over time.

Asked about safety, Malhotra says Hermes does not override the safeguards imposed by model providers and discusses the tension between open access and misuse. His statements describe his position and the product's intended behavior; the interview is not an independent security assessment.

On business model, he says the open harness can be used without paying Nous Research. He describes paid convenience through a model portal and tool gateway, as well as potential business support, customization and private model training. Availability and pricing should be checked against current offerings.

Malhotra describes coordinating people and agents through a Kanban board and discusses proactive assistants. In a flight-booking example he notes that people have different comfort levels and may want approval before actions. The idea depends on the user's instructions, integrations and permissions.

For a personal example, he demonstrates a Sonic Adventure 2 Chao Garden mod. He says he asked Hermes to bring a shrine from another Sonic game into the garden, add animations and a caretaker character, and make game interactions work. The video shows the resulting environment and a Chao interacting with the scene.

Malhotra says members of a modding community praised the work and characterized it as unusually difficult. That ranking is his account of informal feedback, not a published benchmark; the demo also builds on existing games and modding tools.

He recounts the early days of Nous Research, from volunteer model experiments to an open research group and later organization. He credits Technium and the wider team for work on Hermes models and the agent, and describes an earlier project called Forge as a precursor.

Malhotra says Hermes Agent is a major contributor to its own repository. He uses that as an example of its self-improvement workflow, although the interview does not independently quantify the contribution or establish autonomous improvement without human direction.

His stated aim is broad access to capable agents and freedom to choose models, with personalization for each user. That is a vision and commercial positioning from a co-founder; the practical experience will depend on the chosen model, tools, permissions and ongoing human review.