Shadow run: watching before gating
A shadow run is a proposed first pilot: observe agreed activity records without changing existing agent decisions. It shows which recorded actions would require approval under a selected policy.
Why not start with the gates?
Because a gate placed without data either asks about everything or about nothing. The first wastes the owner's attention, the second manufactures the appearance of oversight. Both end with the gate switched off within a week.
Observation gives you the distribution: which actions happen ten times a day, and which happen once a month and cannot be undone. Only then can the tiers be set sensibly.
What actually happens?
We observe activity records within the agreed scope. A shadow run does not block mistakes: your existing controls remain the protection.
- Agent actions are classified against the proposed policy, but never held.
- A ledger builds up: what happened, how it would have been classified, and what would have needed a signature.
- At the end you get the list of actions that gating would have queued, and an estimate of how many decisions a day that means.
- Only then do you decide whether to turn anything on.
What access does it need?
A record of what your existing agents do, in a scope agreed before the start. A shadow run does not require handing over control or changing the tools you already use.
Data scope, retention and who can read it are settled before the first run, not after it.
Where is Gorgos today?
Gorgos is an internal alpha developed for selected KOD.AI workflows. Observing one partner workflow requires source and integration assessment. This is not a ready-made product for self-service activation.
Before a pilot, agree duration, cost, sources, retention and a decision owner. Acceptance covers recorded actions, checked classification examples and estimated owner review effort.