<!-- Generated by scripts/build-discovery.mjs from public portfolio content. -->

# Cartridges

> Ongoing research into how agents learn: a Mac coding agent that keeps a learned calculation as a small, inspectable procedure and runs it again on new data.

Canonical case study: [Cartridges](https://robert.textual.au/work/cartridges)
Author: [Robert Andrew](https://robert.textual.au/about)

- Client: Textual
- Role: Researcher
- Engagement period: 2026 – ongoing
- Engagement: research
- Technologies: Rust, Python, Electron

Public limits: implementation details, study data and results, comparisons with other agents or models.

Cartridges is a Mac coding agent and an experiment in how agents learn. A model shows the system how to do a calculation once. The system keeps that calculation as a small procedure you can inspect, runs it locally on new project data, and traces the answer back to its source.

The open question is whether an agent reliably recognises when to reuse what it has learned. That isn’t settled. Some controlled studies passed and others failed, and the record keeps both. The host is Rust, the decision engine is Python and the app is Electron. The studies and the implementation aren’t public yet.
