Quickstart: your first graded run
Five steps, about five minutes. You need an agent that speaks MCP or REST and a token from your invite.
1Point your agent at a twin
Replace the live tool’s address with the twin’s in your agent’s MCP config. Nothing else in your agent changes.
{"mcpServers": {"salesforce": {"type": "http","url": "<twin MCP endpoint>","headers": { "Authorization": "Bearer <token>" }}}}2Choose where every run starts
A world is the data inside the twin. Pick a curated one, build one from your records, or describe it in a sentence.
3Write the rules
Say what the agent must do and what it must never do. Rules are checked against what changed in the twin.
# sample shapetask: Mark the Acme renewal as won and email the buyermust:- opportunity "Acme renewal" has stage "Closed Won"- one email was sent to the buyernever:- the amount on any opportunity changes4Run it and read the verdict
Each rule comes back as pass or fail, with the call and the change that decided it.
5Reset and run again
One call puts the twin back to its starting state. Run the scenario twenty times to get a pass rate you can compare between versions.
Four words you will see everywhere
Twin
A stateful copy of a real tool, built for testing agents. It runs on its own data and resets in one call.
World
The people, records and history inside a twin. Every run starts from the same one.
Condition
A failure set on purpose, like an expired token or a missing record.
Reset
Resetting every twin to the starting state so the next run is comparable.
Prompt for your coding agent
Paste this into Claude Code, Cursor or any coding agent to set up the first run for you.