cs.CLOct 30, 2026

Fidelity Collapses with Table Size, Not Turn Count: Delivery Architecture in a Deployed Tool-Using Agent

Authors: Tung Thanh Hoang

Organizations: AwardWise, San Jose, USA

Abstract

When a multi-turn agent delivers a tool's structured records into the model's context and re-emits them each turn, every turn can corrupt data the tool computed correctly. We measure this in a production award-travel agent: two architectures over identical frozen payloads, lineage-matched prompts and sampling, differing only in delivery architecture. SYNTHESIS feeds full record sets into the planner's context and answers follow-ups in-head; BYPASS, the deployed system, keeps records in session state, translates follow-ups into arguments for a deterministic filter, and delivers rows over a typed side channel. Across 24 dialogues (339 scored turns per condition), BYPASS scores 92.3% of turns fully correct, SYNTHESIS 45.7% (paired gap +46.6pp [+35.7, +57.1]). Varying delivered-table size by capture date, SYNTHESIS collapses from near-perfect at 10 rows or fewer to 8% or less beyond about 74 rows. On a content-controlled prefix ladder the collapse threshold orders by model capability, from under 10 rows (GPT-4o-mini) to about 198 or more (Gemini 2.5 Pro), while BYPASS stays size-flat for all four models. The collapse is expensive: beyond 100 rows, where almost no SYNTHESIS answer is correct, each correct one costs roughly 200x BYPASS's flat $0.012. Practitioners converged on typed side-channel delivery for rendering, not correctness; we show it was, unknowingly, also a correctness decision.

Figures & tables

Appendix figures & tables12 assets

Supplementary material from the paper’s appendix.

Appendix

Explore similar work

Oct 26, 2026

Context Is a Pipeline, Not a Window: Information Attrition in LLM Agent Harnesses

LLM agents are judged by two numbers: the size of the context window, and how often the final answer is correct. Both numbers miss a third thing: how much of a tool's output is lost before the model ever sees it. We measured this loss in a real tool-using agent. The agent finds the cause of injected faults by reading Linux kernel traces. The context window was almost empty: across 720 runs, the single largest prompt used only 5.75% of it. Yet a third of all tool results (3,827 of 11,189) were cut before they reached the model, and every cut result became broken JSON. The harness cut each result by slicing its JSON text at a fixed length. This does not shorten a result evenly. It deletes whatever data sits near the end. One field reached the model in only 4.7% of calls. A second field, in the very same results, reached the model in 99.3% of calls. Which field held the fault's evidence depended on when the faulty process started: early or late in the trace. The agent's diagnosis accuracy matched this same split. So we ran controlled experiments to test one question: did the data loss cause the drop in accuracy? This paper makes four contributions. First, we model an agent's context as a five-stage pipeline, with a measured failure at each stage. Second, we replay 864 real tool results, including one from a production trace where the shipped cut deletes all 128 records of the fault's evidence. Third, a controlled experiment shows that the cut's real cost appears first as extra work, not as wrong answers: 43% more tool calls, twice the computation, and yield falling from 0.61 to 0.31. The agent still finds the right answer when it has another way to reach it. Fourth, a survey of seven agent frameworks finds two that cut tool output the same unsafe way. Whether the data cut truly causes the accuracy drop is still an open question.
Apr 1, 2026

Beyond Task Completion: A Verification-vs.-Conformance Gap in Tool-Evolving Agents

Agents that synthesize their own tools ship a second artifact alongside each answer: a software library that future tasks reuse, compose, and depend on. Task completion (TC) certifies the answer; it does not certify the library. On a Claude Haiku 4.5 pilot, we patch the harness to preserve per-tool source and replay every synthesised tool against a held-out conformance suite. Across 222 preserved tools and three protocols, 96.8% record per-tool correctness C=0.00: two protocols silent-rot at 100%, one at 91.7%. Hand-written reference implementations score C=1.00 on all 16 capability suites. The synthesised tools execute cleanly and the in-session verifier raises no flag; on held-out inputs they simply return wrong answers. This is the verification-vs.-conformance gap a pass-rate benchmark cannot see. EvolveTool-Bench makes that gap measurable. Sessions are structured into seed, gap, variant, composition, regression, and adversarial roles; runs emit per-tool manifests, verified-subset TC, correct-vs.-incorrect reuse, and audit traces for post-deployment monitoring. In the Haiku pilot over five protocols (3 seeds, 8 sessions, 51 verified decisions per pass), TC alone does not separate the protocols; the audit layer does. One pre-specified contrast shows the adapted decision-split synthesis protocol underperforming one-shot synthesis (-7.1 pp, 95% CI [-13.9,-0.1]; not significant after BH correction) -- a hypothesis-generating signal, not a confirmed effect. We release the harness, per-tool source preservation, conformance replay, and reproducibility manifests. An audit schema of this shape -- verifier coverage, reuse decomposition, held-out conformance -- is the minimum a deployed tool-evolving agent should emit between releases.
Nov 11, 2026

ToolTurns: Benchmarking LLM Agents in Multi-Turn Dialogue

Tool-using language models are increasingly deployed in conversational settings, yet most large-scale tool-use benchmarks focus on isolated requests. We introduce ToolTurns, a benchmark of 2,620 problems spanning 1,827 tools and 11,004 API endpoints across 47 categories. Each dialogue begins with an intentionally underspecified request and continues through two context-dependent follow-ups, testing clarification, tool use, and retention of conversational context. We also introduce an open-source user simulator that reveals withheld context through dialogue and generates follow-up requests, with a checker--fixer pipeline for improving its turns. Across nine language models, the best agent resolves 69.6% of individual requests but completes only 39.8% of full dialogues. Under this evaluation protocol, every evaluated model receives lower success scores on follow-up requests than on the initial request, and dialogue-quality rubric scores and task success can rank models very differently. Paired turn-level evaluation shows that the checker--fixer reduces detected dialogue defects, and blind human evaluation favors the revised turns. ToolTurns provides a testbed for evaluating whether agents sustain task success across multi-turn dialogue with tools.