The gap
AI agents produce two things. We only keep one.
An agent builds software from a plan, architecture decisions, and context — then writes the code. Only the code is saved in git. The reasoning evaporates when the session ends.
persisted
Code → git ✓
The final artifact survives — but not why it looks the way it does.
lost
Reasoning → gone ✗
Plans, alternatives considered, the approach that finally worked — all gone at session end.
Reasoning is lost
Months later you reverse-engineer intent from code. Why this architecture? What was tried first? No record.
Agents work blind
Multiple agents on one repo can't see each other's intent — so they duplicate work and make contradictory changes.
And it compounds: agents run 10–100× faster than humans, and with N agents the conflict space grows as O(N²).