SofiaOS · KarpeOS · AI-SDLC · SOFIA research[email protected]
Independent AI R&D + Product Lab

Operating systems for AI agents

SofiaLabs makes software where agents run a company's real process from start to finish, and a person approves the result. Every step is recorded, so you can check how it got there.

  • A person approves every result
  • Every step kept in git
  • We use it before we sell it
Evening in the lab: a quadruped robot and a caged indoor drone on the floor, and a person at the desk watching two wall screens, the Swarm C2 console with a simulated warehouse incident and the SofiaOS agent board.
Agents and robots under one supervisor. Scene illustrated; on the screens, our Swarm C2 console (simulated data) and the SofiaOS board.

What we build

One platform, and the work around it

Research gives us the ideas and AI-SDLC gives us the method. Both go into SofiaOS, and KarpeOS is the first product built on top of it.

Platform

SofiaOS

A desktop workspace where a company's process runs as a team of agents. Work passes from role to role, and a person approves the result.

  • Decision gates between roles
  • Automatic quality checks
  • A clear view for approval
  • Git history you can audit
Visit sofiaos.com
A tidy joinery workshop: a half-built plywood cabinet, its technical drawing and a laptop with a chat window on the oak workbench.
Built on SofiaOS

KarpeOS and KarpeAI

SofiaOS set up for joinery and woodworking shops. KarpeOS is the system, and KarpeAI is the assistant you talk to inside it.

In development
A SofiaOS delegations board: tasks move from queued to working, review, owner review and done, each with its agent, branch and commit.
Method

AI-SDLC

Our playbook for taking software from idea to production with coding agents: roles, guardrails, checklists for each phase, and evals that catch regressions.

Internal playbook
Our Swarm C2 console on simulated data: a vision-language model explains why an unattended backpack needs a closer look, and a ten-metre cordon is drawn around it while robots are dispatched.
Research

SOFIA

Research into agents that reason slowly and on purpose. When one reflects on a mistake, it updates what it knows about the world, not just the wording of its next prompt.

Long-term work, whitepaper in progress

How we work

We use it ourselves before we sell it

  1. Build it for ourselves

    Every tool starts as something we need in our own daily work. If we stop using it, we stop building it.

  2. Prove it on one real process

    One company and one process, run from start to finish. It has to beat doing the same job by hand.

  3. Turn it into a product

    Only what survived real use gets packaged so that other people can run it without us.

Principles

What stays the same in everything we ship

A person approves
Agents prepare the work. Someone accountable reads a clear, non-technical view and signs off or sends it back.
Every result can be audited
Inputs, decisions and outputs are versioned in git, so you can see who changed what, and when.
Agents say when they are unsure
When an agent is blocked or lacks information, it stops and says so. It does not guess.
No claim without proof
We show software running on a real process. You will not find invented numbers or borrowed logos here.
Contact

Have a process that should run itself?

Tell us what it is and who does it today. We read every message.

Contact us