# Cursor's Agent SDK turns an agent into a folder you can diff. Test what it does, not what it says.

*2026-10-28 ⬩ 8:00 AM ET | bryn*

Cursor made its Agent SDK generally available on October 20, and it turns an agent into a folder you can review like any other code. In [Cursor's changelog](https://cursor.com/changelog/agent-sdk), the prompt is one Markdown file, `instructions.md`. Each file in `tools/` is a single typed action the model can call.

The part I'd copy is the eval. It sends the agent a fixed message and asserts on what it did: which tools ran, with what input, and the shape of the reply. That's a test of behavior, not of prose. Each conversation also keeps a full record of every message and tool call.

Here's what I'd do with it. Write the first eval for the action you'd regret, not the answer you'd like. If the agent can send an email or change a record, assert that an off-topic message triggers neither. Then every prompt edit becomes a diff a teammate reads, and a test that can fail before it ships.

A prompt in a settings box has no history. A prompt in a folder has a blame line.

## Sources

1. [Agent SDK](https://cursor.com/changelog/agent-sdk). Cursor, Oct 20, 2026.

See also: https://www.civic.com/field-notes/version-controlling-your-go-to-market

Source: https://www.civic.com/bin/cursor-agent-sdk-test-actions