Every agent is a pet
Hand-configured, stateful and fragile. You can't reproduce one, you can't say what it will touch, and when it breaks there is nothing to roll back to.
Independent research lab · Barcelona · est. 2026
We build operating systems for autonomous software.
Agents today are pets: hand-configured, fragile, impossible to reproduce. We think they should be infrastructure: declared, versioned, instantiated on demand. We build the machine that treats them that way.
01Premise
Every team running agents today rebuilds the same fragile scaffolding by hand. Agents belong in the same category as the rest of your infrastructure, and putting them there is a systems problem nobody has finished solving.
Hand-configured, stateful and fragile. You can't reproduce one, you can't say what it will touch, and when it breaks there is nothing to roll back to.
Guardrails, approval prompts, tool allow-lists, all written in the same process the agent controls. A boundary only counts if the thing inside can't reach it.
Identity, isolation, supervision, an audit trail. Proven for decades, never assembled into a system an autonomous process can run on. That gap is the opportunity.
02Primitives
Each one is settled at the systems layer, with decades of production use behind it. None of them were designed for software that acts on its own, and making them hold for that workload is the work.
Agents get a system account of their own, separate from yours.
a userIts tools are a declared, reproducible set. Nothing else exists.
a build closureWhat it may read, write and reach is enforced by the kernel.
a sandboxEach task is a supervised process, created and removed on demand.
a transient unitDurable state, plus an append-only record of everything that ran.
a store and a journalThe model is a component inside it, not the architecture.
a runtimeThe mechanisms are proven. What does not exist yet is the system that holds them together for software that acts on its own, and that is what we are building.
03The flagship
An operating system for autonomous software.
Praxis is the first system we've built on those primitives. You declare what a kind of agent is; the machine gives every one of them a real identity, a fixed set of tools and a boundary it cannot argue its way out of.
Keeping these four apart is the whole design.
The control surface is commands, not conversation.
They exist when nothing is running. An agent is state on disk, not a live process.
An agent's toolset is declared up front. What you didn't give it is absent.
Isolation comes from the kernel, so a boundary holds whatever the model tries.
One command prints exactly what an agent may do, and what enforces each line.
04The lab
Drac Labs is an independent research lab working on how autonomous software is identified, contained, observed and stopped.
Everything we learn goes into what we ship, and everything that breaks goes back into the research. It's a short loop, and it's the whole method.
Identity, isolation and audit either exist in the design or they never arrive. So that is where we start.
A tool that overstates its guarantees is worse than one that promises nothing, because you'll trust it.
Anything you can talk your way past isn't a boundary. We work where the model has no vote.
Operating systems are measured in decades. That horizon decides what we build and what we leave alone.
05Draco
Draco winds between the Bears and never sets: a circumpolar constellation that has been overhead every night of recorded history. Four and a half thousand years ago one of its stars, Thuban, was the pole star: the fixed point everything else turned around.
We took the name because it is the right metaphor for the work. Something enormous, old, and useful precisely because you can navigate by it, for as long as you keep watching it.
The figure drawn in the background is the real thing: seventeen stars at their true right ascension and declination, stereographically projected around the asterism's centroid.
06Contact
We're early, and we'd rather talk to a handful of the right people than announce anything. If one of these is you, we should speak.
Deep tech and infrastructure, with a decade-long view of where agents will run.
Teams already running autonomous software somewhere it actually matters.
Systems people who want to work where the guarantees are actually enforced.