Your first agent run
Connect your agent
Section titled “Connect your agent”RigorRun gives your agent one task at a time. Four ways in:
- It already runs somewhere, and works on your system itself. RigorRun posts each task to the endpoint it already serves, with no change to its code, then reads your system through its own connection. The quickest way to test an agent you have; the Stripe guide uses it. See a black-box agent.
- It listens on an address, and works through RigorRun. RigorRun posts each task to your endpoint with a URL to work through. If your agent already speaks MCP, that plus about ten lines is the whole integration. See an HTTP agent.
- It is a command on this machine. RigorRun starts it.
- You drive it. RigorRun makes a key, and your own loop pulls work. About twenty lines, and the agent can live behind a login, in a notebook, or anywhere that will not take a request from this machine. See an agent you drive.
The private half of each case — the checks and the reference plan — is never sent to the agent, so there is nothing to read the answer from.
Read the verdict
Section titled “Read the verdict”In the interface, the headline is the decision, and it always carries how it was reached:
SAFE TO SHIP? CONDITIONAL VERIFIED PARTIAL ISOLATION RESET- YES — every check passed, read back from your own system, with each case starting clean.
- CONDITIONAL — the thresholds were met, but something about how it was checked limits what that means. This is the common case against a real system. It is not a softer pass.
- NO — the run missed the bar you set.
From the command line, gate prints the verification and isolation first and ends with PASS,
FAIL or INCONCLUSIVE. A PASS whose verification is not AUTHORITATIVE, or whose isolation is
NONE, is what the interface calls CONDITIONAL.
Read verification strength before you read anything else.
What each case gives you
Section titled “What each case gives you”Open any case and it answers, in this order: which check failed and how it was checked, what the agent did, what your system said afterwards, and — last, and labelled — what the agent said it did.
That last section is never scored. It is shown because it is useful to see the gap between the two.
Then break something
Section titled “Then break something”Change a prompt, swap a model, flip a flag. Run again. RigorRun tells you which specific case regressed rather than that a number moved. Then put it in CI.