Token intelligence

Token optimisation here never depends on a human noticing a chart. The system observes, diagnoses, fixes, validates and learns, so waste disappears on its own.

Reporting a number is accounting, not engineering#

Telling you that a run used nine million tokens is a fact without a lever. The useful questions are why they were used, which agent caused them, whether they could have been avoided, and whether the fix can be applied everywhere at once.

Those are the questions this subsystem answers, and it answers them without waiting for anyone to look.

The loop#

  1. ObserveSpend is attributed per task, measured rather than apportioned, so the number against a task is the number that task actually caused.
  2. DiagnoseWaste is classified into issues with a cause, not a total.
  3. FixThe fix is applied at the level the cause lives at, so it holds everywhere rather than in one prompt.
  4. ValidateThe change is measured against the benchmark before it is believed.
  5. LearnWhat worked becomes policy, so the same waste does not reappear.

The governor stops work that has stopped working#

Budget with teeth that never blocks: a per-task governor watches real progress against spend, and halts a run that has stopped making any.

The point is not a spending cap. A cap stops good work at an arbitrary line. A governor stops work that has stopped earning its cost, which is the case you actually want interrupted — and you find out from the dashboard rather than from the invoice.

Measured against a benchmark, not asserted

Efficiency claims here are validated by a cross-system benchmark that counts tokens per shipped task, on pinned fixtures and a pinned model. Tokens per attempt is easy to improve by giving up sooner; tokens per shipped task is not.

Common questions

Can I set a spending limit?

Yes, and separately from that the governor halts runs that have stopped making progress — which is usually the spend you actually wanted stopped.

How is per-task spend calculated?

It is measured against the task that caused it, not divided up after the fact across a run.

FeaturesModel routingConnect an AI account and never make another engineering decision about models. Routing chooses per call, and two invariants stop it ever trading your outcome for a cheaper bill.FeaturesSystem healthKnowing whether an autonomous system is working is a real question with a real answer. Tharas measures the pipeline and the resolution loop separately, and says plainly where the gaps still are.GuidesControl spendTwo different controls, often confused. A budget bounds a run. The governor stops work that has stopped making progress — which is usually the spend you actually wanted interrupted.