The agent framework is becoming the new lock-in

Hermes Agent picked up X Premium and Grok subscriptions in 48 hours while NVIDIA pitched it as the on-device runtime. Singapore plans for 1.3B agents. The framework, not the model, is where lock-in lives now.

This week made it concrete: the agent framework, not the model, is becoming the integration substrate. Hermes Agent — the orchestration framework from Nous Research — picked up X Premium and Grok subscriptions as first-class integrations within two days of each other, while NVIDIA positioned it as the on-device runtime for RTX PCs and DGX Spark. The same week, Singapore's head of AI Govtech told swyx the country is planning for 1.3 billion agents within two years and building a national MCP gateway to route them.

The Hermes consolidation is the clearer signal. On May 15, 2026, xAI enabled Grok subscriptions inside Hermes Agent. The next day, X Premium followed — with Hermes now able to search X posts directly. NVIDIA's RTX AI Garage post frames the value plainly: identical models perform better inside Hermes than in thin-wrapper alternatives, because the framework runs as a persistent orchestration layer rather than a task-by-task executor. The pitch is no longer "use our model" — it's "use our agent runtime, plug in whatever models and subscriptions you want."

The Singapore number reframes scale. 1.3 billion agents for a country of roughly 6 million people implies hundreds of agents per resident, most of them ambient and routing through a public gateway. That is an MCP design assumption — protocol-level interop matters when the agent count exceeds the human count by two orders of magnitude.

For working developers, the implication shifts where you spend evaluation time. The model layer is increasingly fungible — simonw noted this week that coding agents have collapsed the cost of porting a native mobile app to React Native, then porting it back if you change your mind. Programming languages, frameworks, and even runtimes are losing their lock-in. What does have lock-in is the orchestration layer: which subscriptions it can authenticate against, which tools it can call, which MCP servers it speaks to. Comparisons like Second Talent's CrewAI vs AutoGen breakdown start to matter more than which underlying LLM you picked.

The bottom line: Pick your agent framework before you pick your model — in 2026, the framework is where the integrations, the subscriptions, and the practical lock-in live.

Sources