Vorza

Preparing software for its next users: AI agents

[Vorza @useVorza]

MCP or CLI: which surface do agents actually succeed on?

We ran the experiment. Stripe, 6 identical tasks, both surfaces, Claude Code and Cursor, 4 runs per cell, fresh sandbox every run, every message recorded on the wire.

[Vorza @useVorza]

@_rafikkk Stay tuned!

[Vorza @useVorza]

MCP or CLI: which surface do agents actually succeed on?

We ran the experiment. Stripe, 6 identical tasks, both surfaces, Claude Code and Cursor, 4 runs per cell, fresh sandbox every run, every message recorded on the wire.

[Vorza @useVorza]

@hasanwkhan04 Stay tuned for Vorza Verdict 002!

[Vorza @useVorza]

Vorza Verdicts #002 next week.

Want this run on your MCP, SDK, or CLI, including your authenticated surfaces? https://t.co/TqRwRoecQL

6/6

[Vorza @useVorza]

Scope: these 6 tasks, these versions (in the image footer), 4 reps per cell, Wilson 95% CIs, test mode. Cursor cannot be tool-restricted, so its MCP cells measure whether it chooses and succeeds on the surface. Full methodology: https://t.co/gu8LQpI31y

5/6