Two things break multi-session agent work, and both are about state.
The first is amnesia. Every new session starts blank, so without a deliberate record you re-explain context, re-derive decisions, and quietly re-litigate things you already settled. The fix is a handoff at the end of each session — what shipped, what's open, what's decided, what's explicitly abandoned — plus a state audit at the start that pulls live state rather than trusting the handoff's snapshot. Carry the abandoned list forward or you will resurrect a dead idea within two sessions.
The second is collision. Two sessions working the same project will each compose work against a tree the other just changed, and agents adapt silently instead of halting. The defence is knowing the current state before you send, and writing briefs anchored to quoted content rather than line numbers, so a collision produces a halt.
Run the seed: write the handoff.