The 10.6% That Belongs To The Harness, Not The Model

The number is 10.6 percent, and it is a scaffolding number. Researchers at Tsinghua University report that an ACID-compliant data agent improves on state-of-the-art agents including Claude Code by that margin on widely used benchmarks ^[4][2]; the authors' own summary on the paper page puts the comparison on KramaBench ^[3]. Nothing in the claim rests on a better base model. The delta is between agent systems, so it is a claim about the harness ^[14].

What the harness does is specific. The paper reinterprets atomicity, consistency, isolation and durability as semantic guarantees ^[1], implemented as commit-or-retry exploration-execution-validation cycles, transactional skill hubs, confidence divergence-based validation, semantic dependency-aware isolation, and transaction-aware state management ^[5]. The failure it targets is the one operators actually see: partial execution that propagates semantically invalid decisions or leaves the workspace inconsistent ^[6].

The reason to hold the figure loosely is the missing second axis. A study from ServiceNow AI Research and collaborators makes the point that memory, workflow and skill modules consume test-time tokens, and that this cost is rarely reported next to the actor's inference cost ^[7]. When they budget-matched, a vanilla actor given a longer interaction horizon instead of the default 10 steps ^[11] matched or surpassed Agent Workflow Memory, Agent Skill Induction and ReasoningBank in aggregate success across three WebArena domains and three models, often on fewer total tokens ^[8], with the same pattern on WorkArena-L1 with Qwen 3.6-27B ^[9]. They also argue run-to-run variance materially affects outcomes and belongs in the evaluation criteria ^[10].

The supporting literature is not kinder to unmeasured scaffolding. AgentCL finds naive task streams barely distinguish memory designs at all, and that naive and held-out settings can expose memory-induced degradation ^[12]. Separate work describes an Echo Gap in which self-graded episodes receive inflated scores and the agent preferentially reuses the mistakes it is most confident about ^[13].

What would move the 10.6 percent: a token-matched rerun against a plain actor with a longer horizon, variance across seeds, and artifacts. The Hugging Face page currently lists no models, datasets or Spaces citing the paper ^[15].