Loading page
The decisions, checks and failure modes we design for, written down by the engineers who build them. Useful whether or not you ever work with us.
Every note so far, newest first.
An agent that cannot say "not this one" will improvise instead. We design the refusal before the happy path, give it a type, and test it as hard as the answers.
5 min read
Flipping the page with dir="rtl" is the easy part. Mixed-direction messages, streaming tokens, digits and icons are where chat interfaces go wrong, and most of it is fixable in markup and CSS.
5 min read
India and Saudi Arabia both route invoices through a government platform, but in a different order, in different formats and with different failure modes. What a pipeline that handles both has to verify, and when.
5 min read
The evaluation decides whether an agent ships, so we write it first. How we build one: where the cases come from, how they are graded, and what the bar is agreed against.
4 min read
We do not run a newsletter, so there is nothing to subscribe to and no form that quietly keeps your address. Tell us what you would like covered next, or ask a question about one of these notes, and an engineer will reply.
Bring one process you think an agent could run.
We'll tell you straight whether it's worth building — and what it would cost if it is.