Arguments, with sources
Longer-form writing on the things that decide whether an inference integration survives contact with production: failover, payment, benchmarking honesty and model churn.
Everything else.
6 posts
- 02
Prepaid credits are a bad deal for agents
Every gateway sells you a balance before it sells you inference. That float is non-transferable, custodial, and structurally hostile to software that decides its own spending.
Economics4 min · Nadia Reinholt - 03
Measuring gateway overhead honestly
Every router advertises a single-digit millisecond number. Most of them are measuring something that does not include the parts that hurt. Here is the methodology we hold ourselves to.
Engineering5 min · Tomas Bergström - 04
What breaks when a model is retired
A deprecation notice is a bug report with a deadline attached. An inventory of everything that silently depends on a model version, and how aliases turn a fire drill into a config change.
Operations5 min · Priya Raghunathan - 05
Why we settle in USDG on Robinhood Chain
Choosing a settlement rail is an engineering decision with a regulatory shadow. The case for a Paxos-issued dollar on an Arbitrum Orbit L2 — including the parts that are trade-offs.
Settlement5 min · Oskar Lindqvist - 06
Building an agent with a spend cap
A walkthrough of giving autonomous software its own key, its own budget and its own audit trail — and why the cap has to live with the signer rather than in an admin console.
Agents5 min · June Watanabe
Everything here is
checkable.
Every figure we publish carries a source and a date, and where sources disagree we print the disagreement instead of picking the flattering read. If you find a number here that has gone stale or a claim that does not survive its citation, tell us — the correction goes on the post, not into a quiet edit.