← All releases

The Loop comes online: classify first, deliver context, then execute

Real-time intelligent context engineering goes from prototype to architecture. The first end-to-end run of The Loop lands mid-week, and the install surface gets sturdy enough to ship across machines.

The cut-line

This is the inflection week. On March 24, the first end-to-end version of The Loop landed — classify, deliver, execute, in that order, on every prompt. Before this week, gramatr was a scaffolding of intent. After it, the architecture was decided.

The publicly verifiable proxy is the contribution-calendar week of Mar 22, 2026 — 599 contributions logged.

What classifying first changes

Until this week, every turn paid the context-rebuilding tax: the model had to figure out what kind of work the turn was while doing it. After the inflection, classification became a separate, lightweight step ahead of the model — and the model receives the work and the context it needs to do that work, nothing more. The savings compound across every turn of every session.

Training the flywheel

Alongside the architecture, the learning pipeline went in: every classification produces a labelled signal the system can learn from. The flywheel that makes the platform sharper with use begins running this week.

Install quality

A multi-client surface landed in the same window, with install hardening across it. The platform stopped being something that ran on a developer’s laptop and started being something that could be installed elsewhere and survive.

← All releases