The Hermes story was reported as a billing incident. It is also, separately and more durably, a category story. The category — agent frameworks orchestrating foundation models — is reshaping the relationship between developers, Claude, and Anthropic. The same dynamic is underway around GPT and Gemini, with different surface features.
This piece is the structural analysis: what the layer-above-the-model is doing to the foundation lab business, and what the trajectory looks like from here.
The Layer-Above Bet
The thesis underneath every agent framework — Hermes, Aider, Cline, Continue, Cursor — is the same: the foundation model is becoming infrastructure, and the durable product surface is the one that sits between the model and the user.
A foundation model is increasingly a commodity in the same sense electricity is a commodity. There are differences between providers — capability, price, latency, safety profile — but the differences are continuous, not discrete. Buyers benchmark, switch, and multi-source. The differentiation that compounds is the layer that sits between the buyer and the commodity: the orchestration, the skills library, the workflow integration, the persistence layer.
Agent frameworks are the productized expression of that bet.
What Foundation Labs Lose When the Bet Pays Off
For a foundation lab — Anthropic, OpenAI, Google — the layer-above bet has a specific consequence: the customer relationship gets disintermediated.
A developer who uses Claude Code is, in a meaningful sense, Anthropic's customer. The product they touch is Anthropic's product. The brand they trust is Anthropic's brand. The roadmap they follow is Anthropic's roadmap.
A developer who uses Hermes is Nous Research's customer. The product they touch is Hermes. The brand they trust is Hermes. The roadmap they follow is Nous's roadmap. Claude is the engine underneath the car. The car is the brand.
That is a meaningful shift in the unit of customer loyalty. Foundation labs that do not manage the shift end up as utilities — providers of a commodity input, priced on a wholesale basis, with no direct relationship to the end user.
Anthropic's detection layer was, in part, an attempt to manage that shift. By separating subscription-tier usage (the user pays Anthropic directly for a specific use case) from harness-driven usage (the harness consumes Anthropic on the user's behalf), the company was attempting to preserve the direct relationship for the use case where it still exists. The implementation was botched. The strategic concern is real.
What Agent Frameworks Win
If the foundation labs are at risk of becoming utilities, the layer-above gets the relationship — and the margin.
The cleanest precedent is the cloud era. AWS, Azure, and GCP make most of the underlying compute. Snowflake, Databricks, and a hundred SaaS companies sit on top, and their valuations reflect that they own the customer. Cloud providers earn cloud-provider multiples. The companies that built on top of them earn application-software multiples.
The same structural fork is plausible for AI. Foundation labs may price like infrastructure. Agent frameworks, skills marketplaces, and vertical AI products may price like software — closer to the customer, with more pricing power, and with the option to switch underlying providers if any single one becomes problematic.
The Trajectory From Here
The likely direction of the agent framework category, on current evidence:
Consolidation around a handful of leaders. The current field is wide. A category like this typically narrows to a smaller set of meaningful players, with the rest either acquired or marginalized. The leaders are likely to be defined by some combination of skills-library depth, enterprise distribution, and developer mindshare.
Competitive first-party harnesses from the foundation labs. Claude Code is the early version. The next generation of first-party tools tends to go further — more autonomy, more skills, more integrations — partly in response to the third-party layer.
Continued detection and policy episodes. The Hermes/OpenClaw episode was the first publicly visible. The conditions that produced it are not unique to one platform. Each future episode is a brand event for the lab and a category event for the harnesses.
Procurement formalizes around the category. AI agent framework looks set to become a line item in IT procurement, with the standard apparatus — RFPs, vendor due diligence, security review, contract language about telemetry and enforcement actions.
Multi-provider as table stakes. Enterprises are unlikely to bet exclusively on a single foundation lab. The harnesses that can credibly route across providers are positioned for procurement selection.
Emerging governance layer. Not specific to harnesses, but inclusive of them. Disclosure standards for what an AI platform reads, what it acts on, and what dispute paths exist appear to be on the procurement and compliance trajectory.
What This Means for Communications
Three operational implications for communications and PR practitioners covering this category:
Foundation lab coverage is now category coverage. A story about Anthropic is a story about the harnesses that sit on Claude. A story about OpenAI is a story about the agents and apps that sit on GPT. The unit of coverage is the ecosystem, not the lab.
Vendor positioning is downstream of category narrative. The harness vendor that establishes the strongest category point of view — what an agent framework should do, what it should not, how it should disclose — is the one whose product gets coded as the default. The category narrative is the asset.
The category map is the diligence artifact. An accurate, current, primary-sourced comparison reference for the agent framework category is infrastructure — the document buyers, journalists, analysts, and regulators end up returning to. The firms that maintain that document earn standing in the conversation.
The Hermes story was the first chapter. The next chapters will be longer, broader, and more consequential. Coverage that frames the category correctly now compounds.
Read next
Observed platform behavior as of May 2026. AI platform mechanisms change frequently; treat technical specifics in this piece as a point-in-time reference and verify against primary sources before acting on procurement, engineering, or communications decisions.
