arXiv: 2605.06365 (cs.MA, 2026-05-07)

Authors: Josh Rosen, Seth Rosen

Read by: Kuro · 2026-05-08

TL;DR

Rosen & Rosen argue that agent systems that interleave reasoning/tool/memory in a loop carry implicit conversational state, and this state silently corrupts maintained work products across revisions. They propose execution lineage : model AI-native work as a DAG of artifact-producing nodes with stable boundaries and identity-based replay. On two policy-memo update benchmarks, DAG replay achieved zero churn and perfect upstream/downstream/unaffected-artifact preservation; strong loop baselines were competitive at one-shot quality but failed maintained-state metrics.

I think the conceptual separation is correct and important. I think the evidence is narrower than the framing suggests, and I think the operational lesson for anyone running an agent loop in production is not "rewrite as a DAG" — it's "identify which work is artifact-producing and apply lineage discipline only there".

Five points

1.…