Agents and Models

We use Cursor’s assistant in two main modes: inline in the editor (online, real-time) and background or task-based (e.g. subagents for parallel doc generation). This page explains how we think about agents and models in this project.

Online vs offline

• Online — The assistant is connected to the Cursor backend and can use the latest model, MCP servers, and your open files. This is the default when you ask for help in the IDE.
• Offline / local — Not a primary mode for us; we assume the assistant has network access for MCP (e.g. Render) and for fetching context. If we add local or air-gapped workflows later, we’ll document them here.

When we use the assistant

• Editing code — Rules and skills guide the assistant so edits respect ports/adapters, tests, and docs.
• Documentation — Skills for generating API/command/reference docs and authoring customer docs. These can be run in parallel by subagents when multiple doc types need updating.
• Repetitive workflows — Commands (workon, open PR, create issue) and scripts (Linear, Render) reduce the need to remember exact invocations.
• Validation — Renovate upgrade validation follows a dedicated skill so the assistant knows how to analyze changelogs and usages before running tests.

We don’t mandate a specific model; Cursor’s default (or your chosen model) is used. The important part is that rules and skills are model-agnostic—they’re just structured instructions that any capable model can follow.

Subagents

For tasks that can be split (e.g. “generate all doc types”), we use subagents so that API docs, command docs, reference docs, and customer docs can be generated in parallel. Each subagent runs the relevant skill. This keeps the main interaction responsive and avoids one long, sequential run.

Extending this

If we introduce new agent types (e.g. review-only, or a dedicated “docs agent”), we’ll document them here and add any new skills or rules to the Reference via pnpm docs:ai-ecosystem:generate.