builder notes — february 2026

february was about getting aura into the hands of a small group of early builders.

we started the month with an internal build that worked about 80% of the time. we ended it with a build that works about 95% of the time. the gap between 95 and 99 is, as always, the interesting part.

the automation runtime got a major stability pass. the main issue was context leaking between automation runs — a workflow that ran in one tab was occasionally influencing the output of a workflow in another. we fixed the root cause and added isolation tests.

the onboarding flow was completely reworked. the original version tried to explain too much upfront. the new version gets you to a working automation in under two minutes and explains the model as you go.

we also shipped the aura updates service, which handles versioned builds, signed installers, and silent updates. it is the plumbing that makes shipping incremental improvements manageable.

the main lesson from february: simple, reliable primitives beat clever abstractions, especially when your users are already juggling complex work.

we had a version of the context engine that was doing a lot of automatic inference. it was impressive in demos. in practice, it made mistakes that were hard to explain and harder to undo. we replaced it with a more explicit model that does less but always does what you expect.

we also learned that the people who give the best feedback are not always the ones who complain the most. the most useful sessions were with builders who just quietly used the tool and told us later what felt wrong.

we are also beginning work on the plugin api, which will let external tools hook into the context and automation layers. more on that when it is closer to ready.