AI agents can reason, plan, and use tools. But doing real work often requires something much more basic: a computer of their own.
Until now, Rival agents and functions executed individual tasks in fast, isolated environments. That works incredibly well for discrete jobs, but sustained work is different. Some tasks require installing software, creating files, running code, making changes, and picking up where you left off. So we gave agents somewhere to work.
The new Rival VM Sandbox lets AI agents and workflows provision temporary, stateful virtual machines on demand. They can install software, run virtually any code, maintain state across multiple steps, and tear the environment down when the job is done.
It's not just somewhere for AI to run. It's somewhere for AI to work.
Why AI Needs More Than a Prompt
Rival's existing execution environment was designed for speed and scale. A Rival Function runs in an isolated environment, completes its task, and disappears. Nothing needs to persist between executions. That's exactly what you want when thousands of independent jobs need to run quickly and in parallel.
But it's not how all work happens.
Imagine an agent needs to:
Download a repository.
Install its dependencies.
Modify several files.
Compile the application.
Run tests.
Fix what failed.
Run the tests again.
Package the final result.
That isn't a single execution. It's a workspace. The agent needs to remember what happened in step one when it gets to step eight.
From Stateless Execution to Stateful Work
The VM Sandbox introduces a temporary stateful execution environment for Rival Functions and AI Agents. An agent can provision a VM, make changes to it, call multiple functions, install software, create files, and continue working in the same environment.
Those changes persist for the life of the VM. When the work is finished, the temporary environment can be destroyed.
Spin It Up. Do the Work. Tear it Down.
Temporary VMs themselves aren't new. Cloud platforms have long allowed developers to quickly provision computing environments. What's different is who is provisioning them. With Rival, an AI agent or workflow can provision the environment itself as part of completing a job.
The infrastructure becomes part of the workflow.
Run Virtually Anything
There's another important implication. Until now, Rival Functions have operated within a defined set of runtimes, including Python, JavaScript through V8, and Lua. That's useful for a huge number of tasks. But the world's software isn't limited to three languages.
A VM changes that.
Instead of asking whether something fits into one of Rival's predefined runtimes, the VM can create the environment the workload requires. Install the language. Install the dependencies. Install the packages. Run the executable.
The environment adapts to the work instead of forcing the work to adapt to the environment.
Bring Existing Software into Rival
This opens another door we're particularly excited about. There is an enormous amount of useful software that already exists outside Rival. Command-line utilities. Open-source projects. Developer tools. Data-processing software. Specialized libraries. Existing executables.
We shouldn't have to rebuild all of it as Rival Functions. With the VM Sandbox, agents can provision an environment, download the software they need, configure it, run it, and incorporate its output into a larger Rival workflow. That means existing software can become part of an AI-powered process.
AI doesn't need to reinvent every tool. It needs to know how to use them.
Give Agents a Place to Work
A lot of the AI industry is focused on making models smarter. That's important. But increasingly, the bottleneck isn't intelligence. It's execution.
An agent may understand exactly what needs to happen and still lack the environment, tools, software, or persistence required to actually complete the job.
The VM Sandbox starts removing those constraints. An agent can have intelligence from an LLM. Tools from the Rival Marketplace. Knowledge from a Knowledge Base. Connections to outside systems. And now, a temporary computer where it can actually do sustained work.
From AI That Answers to AI that Executes
This capability represents something bigger about where Rival is headed. We don't think the future of AI is another chat window. We think AI will increasingly operate software, manage processes, use tools, write and execute code, coordinate workflows, and complete work across systems.
To do that, agents need more than models. They need infrastructure. They need tools. They need memory. And sometimes, they simply need a computer.
Now they have one.