Agent with a filter and side channel gets 92.3% fully correct turns versus 45.7% for one that retells the table

When an agent receives a ready-made table from a tool and rewrites it into its reply on every turn of the dialogue, the model corrupts data that the tool had computed correctly. The authors of a new paper measured this in a production agent for searching award flights bought with miles. They compared two architectures on the same frozen data.

In the SYNTHESIS scheme, the full list of rows enters the model's context, and the model produces the table anew each time. In the BYPASS scheme, which runs in production, the records are stored in session state and the rows go to the interface without passing through the model. The main metric is the share of turns in which every field of every row is correct: 92.

3% for BYPASS and 45.7% for SYNTHESIS. The paired gap between them was 46.

6 percentage points. Errors are driven by the size of the table in context, not by the number of turns.

What was compared and how

The system is called AwardWise. It is a multi-agent assistant built on Google Agent Development Kit (ADK). A planner routes requests, and subagents query award seat search (Seats.aero) and cash fares (SerpAPI). Both architectures ran on Gemini 2.5 Flash at temperature 0.1 and top_p 0.5. The data came from the same frozen vendor responses. Only the delivery differed.

In SYNTHESIS, the standard pattern for ADK, the model handles refinements such as "only departures after 21:00" "in its head" and prints the table again. In BYPASS, it only translates the user's request into arguments for a deterministic filter. The filter operates on the stored set of records, and the rows go to the client as typed events alongside the model's token stream rather than through it. If the model mistranslates the request into arguments, the result is one wrong but real slice of the data. Retelling, by contrast, puts every field of every row at risk on every turn.

The evaluation covered 24 dialogues, with 339 scored turns per condition. The dialogues were built from the sessions of an experienced award-travel practitioner. The ground truth was computed by the production filter evaluator, and checking was done row by row and field by field, with a turn scored zero if the result was wrong. The corpus contains 408 result snapshots across 36 routes, and the largest result set has 510 rows. Across three model families, the same 46.6 pp gap becomes a range of 46.6 to 55.8 pp.

The accuracy cliff by table size

To separate size from the number of turns, the authors ran several experiments. Twenty dialogues on four routes varied by result size. A six-route ladder gave the model the first K rows of the same result set, so the content stayed constant. A Latin square separated size from row position. In every design, SYNTHESIS collapses as the table grows, while BYPASS stays flat. A control table that is byte-identical to the problematic one holds its level even where the retelling scheme scored zero.

The cliff threshold depends on the model. Across four models, it ranges from fewer than 10 rows for GPT-4o-mini to markedly larger values for Gemini 2.5 Pro. BYPASS stays flat on all four (140 of 144 runs). The authors put it this way: model capability buys a threshold, but only architecture provides immunity.

The cost of the cliff and limitations

The cost of a BYPASS turn does not depend on table size and is about $0.012. With SYNTHESIS, tables above 100 rows yield almost no correct answers, and each correct one costs roughly 200 times more than that flat price. Swapping Flash for Gemini 2.5 Pro raises retelling accuracy from 46% to 72%, but a turn becomes about five times more expensive. In the authors' assessment, buying accuracy from a larger model is an expensive substitute for giving up retelling.

The preregistered hypothesis of superlinear error accumulation across dialogue turns was not confirmed. The null model saturates when the per-turn failure probability is already high. However, value tracing showed that failures are not independent: a corrupted output becomes the input of the next turn. BYPASS recomputes the answer from stored state each time, so this kind of error propagation has nowhere to arise.

The limitations are also stated plainly. The gap reflects the combined effect of three differences between the architectures, and the authors did not separate the contribution of each. That would require separate conditions, which they left for future work. The main conditions ran on Gemini 2.5 Flash, inherited from the product's history rather than chosen deliberately. Along the way, the measurements exposed eight defects in the experimental setup, three of them in the authors' own ground truth, plus two corrections to the ground truth's semantics and three production bugs. All were fixed, and all the main numbers were obtained after the fixes.

The takeaway for developers: a typed side channel appeared in LangGraph, AG-UI, Vercel Data Parts and Google A2UI for the sake of predictable rendering. This work shows that it also solves the problem of correctness. If tens or hundreds of rows sit in context, the user should see them without distortion. So it makes sense to store records in state and filter them with deterministic code, rather than ask the model to rewrite them from scratch.