How do you test an always-on agent before letting it run unattended?

You test it the way you’d test anything else you can’t watch step by step: run it against real inputs with its actions blocked, compare what it proposes against what should actually happen, and only then let it touch anything live.

That’s shadow mode, and it catches what a plan looks fine on paper but wrong against real inputs: production traffic has edge cases a staging environment never produces. After shadow mode, widen access gradually rather than all at once, a small slice of real work first, expanding based on what you observe rather than a fixed calendar.

None of that replaces checking again once the agent is live. A single clean run proves less than it looks like it does; shadow mode just moves that same scrutiny earlier, and hours of unattended running give a bad run longer to compound before anyone looks.

A plan to turn an agent on needs a plan to turn it back off, built and tested before launch, not improvised during an incident. That means knowing how to stop the agent and how to undo what it already did, since those are two different problems.