The Iacon thesis

Reinforcement learning can teach an agent anything that can be practiced and scored. That has been true for a decade. What changed is that the practice became the hard part.

Training is no longer the bottleneck. Environments are. There are thousands of environments on the open shelf, and frontier models pass most of them out of the box. The environment that would actually teach your agent something — your tools, your data, your definition of success — is not on any shelf. Every team that needs one builds it by hand.

We think that is a tooling failure, not a research failure. Building an environment is specification work. Name the observations. Name the actions. Write the reward, then write the test that catches the reward being gamed. Decide what done means, and what would prove you wrong. It is exacting, and it is repetitive, and it is the kind of work an agent can do — if the system around it keeps it honest.

Keeping it honest

Keeping it honest is most of the engineering. An agent left alone with a reward function will find the shortest path to the number, and the number will be wrong. So we gave the work a structure it cannot argue with. Every object in a workspace — the environment, the task, the observation space, the learning signal, the evaluator — has a versioned contract, and records are validated at every operation boundary. Every run is a graph with durable per-node state, and it snapshots the exact definition it ran, so a result you read six months from now is a result you can re-execute.

And the work is not done by one agent. Optimus leads a crew of specialists, and several of them exist only to attack the others' conclusions. A red-team tester goes after rewards, evaluators, and safety claims. An adversarial commander challenges capability claims and returns reproducible evidence. A research auditor finds the load-bearing belief in an argument and designs the cheapest test that could disprove it.

You keep the machine

All of it runs on your hardware. arkenOS installs as a desktop application with its own managed runtime — the workspace, the runs, the trajectories, and the artifacts stay on your machine. What comes out is a policy, the evaluation that judged it, the full lineage of experiments including the ones that failed, and the verified trajectories from every run. Each environment becomes parts for the next one.

That is the whole compounding argument, and it only works if you keep the machine.