Oxagile AI · Articles
Notes from the practice
What we have had to work out on real projects, and what it cost to find out.
2026-10-07 · 5 min read
The apocalypse is currently unavailable
Skynet wakes up, decides to end humanity and gets "Instance not available". On GPU queues, and why doing more with less beats waiting for more.
2026-10-05 · 5 min read
A model change is a release
A prompt edit or a provider upgrade changes what the product tells customers. An evaluation suite and a gate in the pipeline turn that change into a release someone signs.
2026-10-04 · 5 min read
An answer you can trace
Most document systems answer plausibly and wrongly. The fixes sit upstream of the model: structure in the index, version and scope filters, citations, and a defined answer for "not found".
2026-10-03 · 5 min read
The cost of an answer is set before the first prompt
Routing by task class, one task per session, navigation over ingestion, a cached prompt opening and a cap on each session decide what an AI feature costs. A budget only reports the result.
2026-10-01 · 5 min read
Characterise legacy before an agent changes it
Old systems carry behaviour nobody wrote down, and some of it is relied on. Characterisation tests record it before the first edit, so an agent can work fast inside a net.
2026-09-30 · 5 min read
An agent writing code in a regulated scope is now in that scope
Change control assumes an identified, authorised person. An agent-authored commit answers none of that unless the trail was built while the work ran.
2026-09-30 · 6 min read
A test suite can only fail against something someone expected
Automation checks conformance to a prediction and holds no opinion about the prediction. That is the whole reason a person still runs the product before anything gets locked.
2026-09-29 · 5 min read
The writer never reviews itself
A session that wrote the code is holding every decision it made, including the wrong ones. Of the defects caught before merge on our own projects, 1 in 5 was found only by a separate review session. The writing session had already passed it.
2026-09-26 · 5 min read
Navigation instead of ingestion
An agent that reads a codebase to find something pays for the reading on every turn. One that queries an index of how things relate builds it once and carries a 400-token answer instead of 10,000.
