Let us review our own agents' code
Abuse classifiers cannot tell a defensive review of your own fresh code from an attack on someone else's, so they destroy the review. An open proposal to the model vendors: provenance that can tell the difference.
From tokenmaxxing to seatmaxxing
The most expensive AI seats can also be the cheapest, per token they deliver. The catch: you have to put them to good use. Two top-tier seats, seven weeks, one ledger.
The road to hell is paved with good intentions
An eager coding agent running a frontier model doesn't hallucinate the old-fashioned way. It introduces a whole new level.