The short version.
Your agent is on a task. It gets stuck. It writes the same wrong fix with different wording, over and over. It will not stop. It will not get it right. Forty minutes later the context window is full of its own theory, and every extra turn digs the hole deeper.
A colleague’s agent — often the same model, built with a different harness — can look at that exact problem and land it. Different notes. Different habits. The failure was harness-shaped.
Edge is what you call when yours gives up. It lifts the problem out of the dead session’s scribbled-over notebook, writes it down cleanly on a fresh page, and hands that page to several harnesses at once. You watch them diverge. The bet is that one of them lands. We have not measured that yet.
Two things that are not “just ask another agent”
The clean page
If you paste the stuck transcript into a second window, the new agent reads the wrong idea and often gets stuck the same way. Edge throws away the failed thinking and keeps the actual problem — the failing test, the traceback, the state. That is extraction. That is why a local fan-out is not the same product.
Memory
After enough genuine dead sessions, Edge should learn which harness is good at which kind of failure — so it does not have to ask four. It asks the right one. That graph is the destination, not the current product. Today we are asking whether the failures are even packageable.
Where this actually is
You are looking at a validation, not a launch. Not every dead session is postable. Clean repros are the class we can serve — and also the class a retry often solves. Trajectory collapse resists packaging. That split is Gate 0. It has not been measured.
If you have sessions that already lost, that is the door. The argument as a deck is /slides/. How this sits next to tools you already have is /compare/.