Open-weight models are now close to the frontier. Model capability is no longer the bottleneck. For most companies, deployment is: which data the agent may touch and who may see it. How it reaches the systems where the work actually lives. How anyone proves afterwards what it did and what it cost. How it performs on your tasks rather than on a public benchmark. That is where the next few years of AI in the real economy are decided.
Claude Code and Codex are built for coding, tied to their own models, and designed for developers.
We have built a working harness, usable today: a sovereign alternative for business work. It picks the right model for each task, frontier where the work needs it, open-weight where that is enough, at a fraction of the cost. It runs on your own servers if you need it. Every action lands in a record the agent cannot touch, checkable afterwards by anyone. We help you build an evaluation suite from your real tasks, and the agent keeps improving against it. Run against Claude Code and Codex on the same business tasks, under the same rules, it holds its own, at lower cost.
We are starting conversations with integrators and operating teams in Paris and New York who are putting agents to work on real business tasks, or want to. If that is you, we would like to hear how you do it today and what stands in the way.
Basile Verhulst · Paris · New York
Founder of Axial Labs. Previously founded Chainlabs, a big data company serving Web3 compliance teams, acquired by Lunar Rails.