What was released
An open model of 550 billion parameters, of which only 55 billion activate for any given request — a committee that keeps specialists on staff without sending all of them into every appointment. The vendor’s stated purpose is “frontier reasoning and orchestration in long-running agentic workflows”: the model is meant to be the one that plans, checks and recovers, while smaller, cheaper models do the routine tool calls.
The engineering is genuinely interesting — a hybrid architecture that handles very long context efficiently, a compact number format, a single checkpoint across three generations of hardware, and a training method in which more than ten specialised teacher models score the student’s attempts. The commercial logic is the familiar one: an open, inspectable, frontier-class model that runs best on the vendor’s silicon, so that the orchestration layer of the agent economy gets built on that hardware by default.
The performance figures — five times the throughput of comparable open models, agentic tasks completed with fewer tokens, costs cut by up to 30 percent — are the vendor’s own benchmark runs. Wait for reproduction before signing anything. The idea underneath does not need the figures to be right.
The problem it names is real
Agents that plan, call tools, spawn sub-agents and carry their working history from turn to turn burn tokens at a rate that grows with every step, and a long task drifts. Anyone who has run an agent on a multi-hour workflow has watched it wander, repeat itself, or confidently complete the wrong task. The cost is doubled: the tokens spent going astray, and the tokens spent on the person who has to notice.
The proposed division of labour — routine work to cheap models, judgment to an expensive one that steps in for planning and recovery — is the correct response, and it is worth noticing that it is not a new idea. It is how organisations already structure decision-making: routine cases handled at the front line under standing rules, exceptions and judgment calls escalated to someone senior, recovery owned by whoever has the authority to unwind a mistake. The supervisor model is that structure expressed in software.
What a supervisor is, and is not
Two boundaries keep this from going wrong.
The supervisor is a control component, not a control. It improves the odds that a long workflow stays on course, and it makes recovery cheaper. It does not replace the check that the system of record enforces on the agent’s output — the reconciliation rule the ledger applies, the validation the CRM runs, the certified sign-off a foundry demands before a design ships. A supervisor that approves a bad action is still a bad action. The sign-off has to live where accountability lives, in the system, and run on every output regardless of which model proposed it.
The supervisor is also an agent, and needs a job description. Which decisions may it take on its own — retry, re-plan, roll back? Which must it escalate to a person — anything irreversible, anything above a value, anything outside the workflow’s scope? What is its token budget, and what happens when the budget is exhausted? A supervisor with unlimited authority and unlimited spend is the most expensive way yet devised to drift off course.
The operating model, without the vendor
Strip out the product and the shape worth building is this:
- Tier the work. Routine, reversible, well-specified steps go to small, fast, cheap models. Planning, exception handling and recovery go to a larger one. Decide the boundary explicitly, per workflow.
- Budget in tokens, govern in outcomes. Cost per completed task — not per token, not per call — is the number to manage, and it should be measured per workflow with a ceiling that stops runaway loops.
- Keep the sign-off in the system. Whatever the supervisor decides, the system of record validates. Failure is a stop, not a note.
- Make the supervisor swappable. Open weights are inspectable and portable, which is welcome. So is the ability to replace this supervisor with the next one on the same evaluation set, without rewriting the workflow.
- Log the reasoning where it can be found. The supervisor’s decisions — why it re-planned, why it escalated — belong in the same audit trail as the actions, tied to the records they affected.
The interns and the adult
The vendor’s framing was that somebody has to supervise the interns. True. But the analogy carries its own warning: in a well-run organisation, the senior person who supervises the interns does not also sign the accounts. That is a separate control, held by a separate function, precisely so that judgment and verification are never the same step. Build the agent room the same way — an adult in the room, and a sign-off at the door.
