DeepInfra has started renting agents by the hour. Pick OpenClaw if you want a web dashboard, Hermes if you would rather stay in a terminal, and it is running in minutes, billed only while it is up. Alongside it they are renting Sandboxes: isolated Linux microVMs with their own kernel, started with one API call, billed by the second and free while stopped.
Three things you are choosing, not one
An agent setup is usually discussed as a single decision. It is three, and they fail differently:
- •The agent software — OpenClaw, Hermes, something else. This is a UX and capability choice.
- •Where the process runs — your laptop, your cluster, or now somebody else’s hourly VM. This is an availability and access choice.
- •Which models it talks to — and what happens when one of them is down, rate-limited or out of credit.
Bundling the first two is exactly what a hosted agent should do, and doing it well is genuinely useful: an agent that only runs while your laptop is open is not an agent, it is a script you babysit. But the third is a separate problem, and it does not get solved by moving the process somewhere else.
Why the sandbox part is the interesting half
Hardware-isolated microVMs billed by the second is the more consequential piece of that announcement, and it gets the smaller headline. The reason is unglamorous: agents that write and run code are the ones that actually do work, and running model-written code on the same machine as your credentials has been the unspoken risk in every local agent setup since this started.
A per-second microVM makes “run this in something I am willing to lose” cheap enough to be the default rather than the careful option. That is a real improvement in how safe these systems are to operate, independent of who you buy it from.
Where routing fits, and where it does not
A hosted agent points at an OpenAI-compatible endpoint. Point it at one provider and you have the agent’s availability plus that provider’s availability multiplied together, which is strictly worse than either. Point it at a router and a provider outage becomes a slower response rather than a dead loop.
That is not a hypothetical. In the last three weeks our own probe found eight providers that could not serve at all — four out of credit, three on dead credentials, one that had retired its API outright and returned HTTP 410. Every one of them was silently broken: nothing alerted, because a provider nobody is currently calling generates no failures to notice.
If your agent had been pinned to one of those, it would have stopped. Not crashed — stopped, mid-task, with an error in a log you were not reading.
# Whoever runs the agent, the endpoint is the same shape.
export OPENAI_BASE_URL="https://antbase.ai/v1"
export OPENAI_API_KEY="$ANTBASE_API_KEY"
export OPENAI_MODEL="ant:coder"Both agents in that announcement already run on Antbase and have done for a while — we have setup guides for each, including running OpenClaw against free models and managing it remotely on Kubernetes. Nothing about a hosted version changes the model layer underneath it.
When hosting it yourself is still right
Hourly billing is excellent for agents that run in bursts and terrible for agents that run continuously — at which point you are renting a VM at a markup. If your agent is up twenty-four hours a day, price it against a box you already own before you move it.
And if the agent needs to reach things inside your network — an internal API, a database, a document store that will never be public — a hosted agent is the wrong shape regardless of price. That is the case our Kubernetes guide exists for.
