Sep 2026 · Writing
We audited our own MCP server. We failed.
Tool schemas cost more tokens than the work they describe. What we measured, and the two patterns we got wrong.
Winged Victory of Samothrace · Louvre · CC0the torso against the wind · 5.6°Research
What agents get done per dollar.
We settle what agents pay for, so we can measure what the money bought: which models finish the task, what they really spent, and where it went. Any evaluation we publish runs on tasks we synthesize. A private evaluation runs on a customer’s own cases and is never published.
Writing
Newest first.
Aug 2026 · Writing
Routing saves money in three places. Here is where it doesn't.
Measured savings with confidence intervals, and the cases where a router will not help you.
May 2026 · Writing
The first quarter of a market whose buyer is software.
Ninety days of paid agent calls, described from the settlement side.
Mar 2026 · Writing
Speed and intelligence are poorly correlated.
A routing benchmark run through a production payment pipeline, and the classifier we built from it.
Live
Instruments that run on the gateway's own settled calls.
Live
Observatory: where a dollar of agent spend goes.
From settled calls, updated continuously.
Live
Every paid call, as it settles.
The public feed of the gateway.
On your own tasks
The same measurement on your workload. Never published.
Service
Verified Eval: your own production tasks, graded by code.
Every candidate model, under a hard budget, priced per completed task.
Service
Token audit: where your agent spend buys nothing, and a policy to stop it.
Measured on your own tasks, at settled prices.