Power & Economics
An autonomous system that pays per token thinks twice about thinking. Praxis is engineered the other way: flat-rate seats, local models, cloud off by default so the marginal cost of thinking approaches zero.
The Reactor Gauge
Today's burn, by provider from the live Engineering station.
Heavy work dispatches to subscription CLI seats rather than metered APIs. A local model on the Mac Studio handles embeddings, routing, and routine roles for free. The cloud API toggle exists, reads zero, and stays off unless the operator flips it for something that genuinely needs it.
Every dispatch decision is itself instrumented by agent-dispatch histograms that record which executor was chosen, why, and what it returned, so the routing table is tuned on evidence, not vibes.
On a record day the system burned 22.7M tokens. The bill was the same as on a quiet one.
The System, by the Numbers
Counted from the live system — board API, source tree, cron registry, graph, fleet monitor.
Measured snapshot · July 2026