In Sarasota, Florida, a company whose pipeline is built around PTSD and chronic-pain therapies quietly changed a dropdown. When a customer spins up a new agent on QwikAgents, they now choose between two runtimes: OpenClaw, or Hermes.
QwikAgents is the wholly owned managed AI agent subsidiary of Silo Pharma (Nasdaq: SILO), a developmental-stage biopharmaceutical company. On September 30 it announced two additions to its platform: support for customer-supplied NVIDIA API keys, which unlocks Nemotron models, and the Hermes agent runtime as a second option alongside OpenClaw. Customers pick the runtime per agent, at creation time.
A runtime is the least glamorous layer in the stack and the one that does the most work per hour. It is the software that reads a customer's instructions, decides which model to call, carries out the assigned task, and keeps the thread of state alive between sessions. Both runtimes on QwikAgents sit on the same managed hosting, persistent memory, model routing and dedicated compute.
The Runtime Becomes a Line Item
Model choice has been a familiar story for two years now: bring your own key, swap the brain, dodge lock-in. NVIDIA becomes the fifth bring-your-own-key provider on the platform, and it arrived quickly after QwikAgents joined the NVIDIA Developer Program in July 2026.
- Anthropic, OpenAI, Google and DeepSeek, the four providers already supported before this update
- NVIDIA, added on September 30, giving customers access to Nemotron models with automated model routing designed to keep agents available
- Both runtime options, OpenClaw and Hermes, sharing the same managed infrastructure underneath
The second addition is the more interesting one, because agent runtimes have until recently been something a team picked once and then lived with. Turning one into a customer toggle is a newer move. "AI models and agent software are both evolving rapidly, and we believe customers should have the flexibility to choose the technologies that best fit their needs rather than be locked into a single provider or runtime," said Eric Weisblum, Silo Pharma's chief executive. He added that the company expects to keep broadening both the models and the runtimes it offers.
That framing only works if runtimes are genuinely swappable, which is a bigger claim than it first looks. An agent runtime owns a lot of work that nobody advertises.
- Reading the instruction and planning the steps
- Choosing and calling the model, including the routing decisions between providers
- Executing tools, shell commands and browser sessions
- Carrying memory, preferences and session history forward into the next run
Which is exactly why the second half of this story matters. Running Hermes yourself is not a dropdown. A deployment write-up published this week walked through what happens when a team decides Hermes, Nous Research's self-improving agent with more than 230,000 GitHub stars, should live somewhere more durable than a laptop. The conclusion was blunt: Kubernetes is the obvious destination and the least documented one. There is no official Helm chart. Nous Research documents install.sh, Docker, Compose and Nix packages, and stops there, leaving three community charts that disagree with each other and none of which is backed by the vendor.
Three constraints do the real damage.
- State has a single writer. Configuration, MEMORY.md, USER.md, session history in a local SQLite database and every learned skill live under HERMES_HOME, which is /opt/data in the official image. Two active pods behind one volume is data corruption, not horizontal scaling, so the pragmatic pattern is one release per tenant with replicas pinned to one.
- The bootstrap wants root. The official image runs s6-overlay, whose /init chowns the volume on first boot before dropping to the hermes user, UID 10000 rather than the 1000 that most published examples assume. A pod spec with runAsNonRoot set to true blocks startup outright, which means a namespace exemption if your cluster enforces the restricted Pod Security Standard.
- There is no HTTP reload hook. MCP servers and skills do hot-reload interactively, but when a ConfigMap update lands on disk nothing tells the process. The open issue asking for admin reload endpoints is still tagged low priority, so teams wire a config checksum to force a rollout instead.
None of that makes the managed route wrong. It makes the trade explicit. Take the runtime from QwikAgents and the volume ownership, the init sequence and the reconciliation gap belong to somebody else. Run it yourself and you keep the state that makes the agent yours, along with the persistent volume, the startup probe that has to survive that first-boot chown, and the graceful shutdown window you will test for the first time during a node upgrade.
The interesting part of September 30 is not that Hermes took a slot away from OpenClaw on a biotech's platform. It is that the layer where an agent actually lives has started to appear as a checkbox instead of an install. Checkboxes arrive before portability does, though. If a customer ever wants to tick the other box, the memory and the skills still have to travel, and nobody has made that part a dropdown yet.
Comments