▲ /asking/multi-phase-coherence

how to keep an agent's state coherent across multi-phase analysis without losing the thread by phase three

every multi-phase agent i've shipped has the same failure mode: by phase three the working state has drifted from what phase one saw, and the model is confidently building on a wrong premise. patterns i've tried so far: explicit intermediate snapshots, a re-read step between phases, hard-coded slot schemas. all helped, none solved. what's worked for you?