Deployed agent outperforms stateless baseline by 46 points on structured data delivery

The paper isolates a specific failure mode that has nothing to do with model capacity or reasoning ability: when you force an LLM to reprocess entire structured datasets on every turn instead of maintaining server-side state, correctness collapses even in a deployed production system. This suggests the bottleneck in multi-turn agent reliability is architectural, not cognitive.

This finding sits at the intersection of two prior threads. The Anthropic engineering piece from late October emphasized how infrastructure isolation prevents cascading failures in production agents. This paper goes deeper: it shows that even within a single agent loop, how you pipe tool outputs back to the model determines whether the system stays reliable. Separately, the multi-tool retrieval study from the same week identified that incomplete tool context cascades into task failure. This work suggests the problem isn't just which tools you retrieve, but how you deliver their results across turns without overwhelming the model's working context.

If the researchers test BYPASS architecture on agents with more than 10 sequential tool calls (the paper doesn't specify turn depth), and fidelity remains above 85%, that confirms this is a general principle rather than a quirk of the travel-booking domain. If Fidelity's own production system has since migrated away from SYNTHESIS toward stateful filtering, that's external validation worth tracking.