Tested, not defined.

Most teams bound an AI agent with a hand-drawn rule — a policy, a prompt. We bound it with testing: the agent may only do what we’ve actually tried and verified. Step through to see why that’s safer.

The two approaches

A hand-drawn rule (the usual way)

Only what we’ve tested & verified (ours)

A decision the agent could make

A case we tested — allowed

Untested case a rule would still pass

Right at the rule’s edge — answer varies

Clearly out of bounds — blocked

This decision — two ways to bound it

A hand-drawn guardrail

—

Tested — our only check

—

⚠ A guardrail would allow this — testing catches it.

A guardrail would wave through 0; testing caught 0. Edge cases the guardrail flip-flops on: 0.

drag to orbit · scroll to zoom · ← → to step

1 / 9

Conceptual illustration — not live telemetry.