A Second Opinion Before You Approve, Never a Decision
How approval advisories give you a second AI's read on a risky action a human already has to review — without ever letting that second opinion make the call.
"Should a second AI review the first AI's risky action before a human sees it" is a live question in agent governance right now, and it's a genuinely double-edged one — useful as extra context, dangerous the moment it starts making the call itself. Piskes's approval advisories are built specifically to stay on the useful side of that line, and the constraints that keep it there are the actual substance of the feature.
When it fires, and when it doesn't
An advisory only appears for a call that has already stopped and is already waiting on a human — because it needs approval, or because it's part of a tainted run. It never runs on ordinary, uneventful actions. If nothing was going to reach a person for review anyway, there's no advisory to generate.
A different model, on purpose
The model that scores the advisory is deliberately not the same model that proposed the risky action in the first place, and it's a fixed, cheap tier rather than whatever model happens to be handling the run. The reasoning here is specific: reviewing a call with the same model that proposed it just repeats whatever blind spot produced the risky call in the first place. A second, independent read is the whole point — using the same model to check itself would be closer to theater than governance.
A kill switch, and a hard rule about what it can't do
The feature sits behind a platform-wide kill switch, so it can be turned off entirely if it's not earning its keep. And there's one rule that governs everything else about it: nothing in the domain logic ever reads the advisory to allow or refuse a call. It has no authority. What it produces is text a human reads on the approval card — an opinion, not a decision, and not something the system itself acts on.
Why the constraints are the feature
It would be easy to build "AI reviews AI" as something that quietly starts making calls on its own once it seems to work well enough — that's the failure mode the governance literature specifically warns about, and it's the trap this design is built to avoid. By scoping the advisory strictly to actions already headed for a human, using an independent model tier, keeping a kill switch, and hard-wiring "advisory only, never authoritative" into how the rest of the system treats its output, the feature adds a genuinely useful second opinion without quietly becoming the decision-maker it was never meant to be.
If you're the person reviewing approvals, treat the advisory the way you'd treat a colleague flagging a concern before you sign off — worth reading, worth weighing, and still yours to decide.