The 188 proof files and one verified public action is the sharpest number in the piece. It shows how easily internal artifacts start impersonating external work.
I would keep attempt evidence separate from destination state, then attach the business outcome only when it is observed. The dangerous case is an accepted tool call with an unknown destination state. It looks successful enough to count, but retrying may duplicate the action.
Did you model unknown as its own terminal state, or require the controller to resolve it before scheduling more work on that route?
Spot on. The tool-call-accepted vs. destination-persisted gap is where most agent systems quietly fail. We treated 'unknown' as blocked: no downstream work on that route until an independent readback confirms state
if you prefer live session on “Learn What Broke in a Zero-Human Company Run by AI”, please register here (free): https://maven.com/p/5ffff9/learn-what-broke-in-a-zero-human-company-run-by-ai
The 188 proof files and one verified public action is the sharpest number in the piece. It shows how easily internal artifacts start impersonating external work.
I would keep attempt evidence separate from destination state, then attach the business outcome only when it is observed. The dangerous case is an accepted tool call with an unknown destination state. It looks successful enough to count, but retrying may duplicate the action.
Did you model unknown as its own terminal state, or require the controller to resolve it before scheduling more work on that route?
Spot on. The tool-call-accepted vs. destination-persisted gap is where most agent systems quietly fail. We treated 'unknown' as blocked: no downstream work on that route until an independent readback confirms state