Nous Research, an independent AI lab, built Hermes as a terminal-based agent harness that sits on top of Anthropic's Claude Code, revealing a strategic bet that the next layer of AI value creation lies in orchestration, skills systems, and multi-agent coordination rather than in foundation models themselves. The choice to build on top of a competitor's coding agent, rather than competing directly with it, signals a broader industry realization: foundation models are becoming commodities, while the harnesses and frameworks that coordinate them are becoming the products that users actually care about.

What Is Nous Research, and Why Does It Matter?

Nous Research sits in a distinct category of AI organizations, alongside groups like Eleuther and Together AI, that take a fundamentally different stance from major foundation labs like OpenAI and Anthropic. Rather than focusing on building larger or more capable foundation models, Nous invests in the infrastructure layer above those models. The lab's center of mass is in agent infrastructure, model post-training, and open-weight model work, areas where foundation labs have been comparatively under-investing relative to the size of the opportunity.

This positioning reflects a core thesis: the differentiation that matters going forward is not inside the foundation model itself. Instead, it lives in the layer above, where orchestration, skills systems, multi-agent coordination, persistent memory, and environment integration happen. Foundation models become utilities. Harnesses, agents, and skills become the products that drive user loyalty and competitive advantage.

How Does Hermes Work as an Agent Harness?

• Terminal-Based Interface: Hermes operates as a command-line tool that users install and invoke to coordinate one or more underlying models, most commonly Anthropic's Claude through Claude Code, to perform multi-step, multi-turn work.
• Curated Skills Library: Hermes ships with a bundled set of best-practices for specific task types, including coding agents, data extraction, and workflow automation, which the harness loads on demand when the task matches.
• Provider Abstraction: The harness can route requests to multiple model providers through a single interface, including Anthropic, OpenAI, and open-weight models served locally, allowing users to configure their preferred model while the harness handles protocol differences.
• Session Orchestration: Long-running, multi-turn work is the harness's native unit, not single prompts; Hermes manages context, tracks progress, and reorganizes conversations as they grow.
• Credential Routing: Hermes can authenticate against underlying model providers through pay-per-token API keys, OAuth where supported, or by reusing credentials already stored by other tools on the same machine.

Why Build a Harness Around a Competitor's Coding Agent?

The answer lies in the gap between what first-party tools optimize for and what power users actually need. Claude Code, like every first-party coding agent, is built to optimize the user experience that Anthropic envisioned. A harness like Hermes is built to enable a different set of experiences: different skill bundles, different orchestration patterns, different model-provider strategies, and different terminal ergonomics that the first-party tool deliberately does not ship.

That space between what the first-party tool offers and what a power user wants is where Nous is building. The choice to build on top of Claude Code is a function of Claude being the strongest model for autonomous coding work in 2026. Were that to change, the harness would route to a different provider. This flexibility is the harness's strategic advantage: it is not locked to any single foundation model.

For Anthropic, the strategic challenge is the inverse. Every harness that becomes the default for a class of users shifts the relationship: the user's loyalty moves to the harness, not the foundation model. Claude becomes infrastructure. The harness becomes the brand. This dynamic explains why Anthropic implemented detection logic to manage how third-party harnesses consume its models, though that implementation was ultimately botched.

What Does This Reveal About AI's Structural Future?

The Hermes story, while not the kind of story Nous Research wanted to be known for, accidentally validates the lab's central thesis: the relationship between the harness and the user is now its own asset class, valuable enough that a foundation lab built detection infrastructure to manage it. This suggests that the industry is entering a phase where control over the user interface and orchestration layer matters more than control over the underlying model.

Nous Research operates adjacent to the foundation labs, not competing with them. While the lab does ship open-weight model work, its center of mass is in the agent and infrastructure layer. The organization is community-oriented, with serious engineering depth, releasing open-source tools, publishing benchmarks, and maintaining an active developer community alongside production infrastructure work.

Whether Nous's bet on the layer-above-the-model pays out is one of the structural questions for the lab's next phase. But the Hermes incident has already demonstrated that this layer is valuable enough to defend, and valuable enough for independent labs to build around. The future of AI may not be determined by who builds the best foundation model, but by who builds the harness that users trust to coordinate them.