Part 4: The Infrastructure

Nov 8, 2025

Posts in this series:

Where We're Going We Need Computers

There is a lot of computation going on in Vivarium:

The AI agents compute things;
The language system runs the compiler tool chain;
The tools compute things like formulas in a spreadsheet or run a model checker;
The world containing all the above computes things like the physics and constraints on agents.

Computation resources are needed for every component. And the infrastructure is that part that connects all this together, getting the things that need to compute to the compute resources they need.

The Current State of Infrastructure

Software has grown up in a manufacturing world, and many aspects of how we build and operate software mimics how we work with physical matter in manufacturing.

But software is different. It's information, and it doesn't need assembly lines, shrink-wrapped boxes, delivery trucks, or store shelves.

Yet, this is really how we still make software. A programmer "produces" some code that goes into a "build pipeline" where it is squished into "packages" that are put on the shelves of some "package repository" so that other things can come along and "deploy" them to "compute resources" (i.e. CPUs, memory, disk drives, networks).

In this current software world, the application represented by the code the programmer wrote most likely will know nothing about the compute resources it will be run on or probably not even anything about the systems used to "test" it.

And the programming language used to write the application is probably very different from the one that is used to describe the compute resources, so even if the programmer was curious and wanted to experiment with having the application know something about how it is run, the separate languages would be a substantial barrier to that sort of integration.

Infrastructure is also What it Does

In Vivarium, the use of language as the interface offers a significant opportunity to rethink how infrastructure and applications work together.

In the case of an agent, it may be running on a separate machine and communicating with another agent on a different machine via a network connection. But it's also possible that the two agents could be running on the same machine, still communicating over a messaging link. Or it's even possible that the two agents could be running in the same process and communicating directly through something like messages-passing in memory.

The ability to seamlessly manage these various scenarios and integrate that with the language machinery supporting the use of language as an interface is the domain of the infrastructure concept in Vivarium.

Telemetry

Another aspect of infrastructure that is built into every part of Vivarium is telemetry.

The OpenTelemetry standard provides a good foundation for making sense of how a Vivarium instance and all its components are operating over time.

Additionally, this very telemetry can be used as an input source of data to the agents and tooling, creating a dynamic feedback system in real time.

Tying it Together

You've now seen a sketch of our idea of an experimentation platform for inventing, discovering, building, testing, and operating a system of machine intelligence.

There's nothing fundamentally new here; that will come in the actual creation of machine intelligence. But these ideas represent a significantly different way of thinking about the construction of a software system.

There is a ton of work to be done. In the next post, I outline an overall initial roadmap and milestones to build Vivarium.