Fetching from the wire…
01
02
03
04
05
06
07
08
Public story · 2026-08-19 · high
Dual access to parsed and native views, layered on that fix, lifted Pass@1 accuracy by up to 12 points, per the paper.
Why now: The paper is a first mention in this outlet's coverage, flagged in the August 19, 2026 briefing.
StagedWorkspace binds an agent's parsed file records, review diffs, and native content to the same version hash, according to a new paper on arXiv (2608.18050).
Agent stacks that hand different components separate copies of a document risk one part acting on stale content while another sees the latest edit. StagedWorkspace measures that failure directly and shows fixing it is worth up to 12 accuracy points on document QA benchmarks.
The paper's fix pairs that version binding with dual access. That lets the agent pull both the parsed view and the native file, instead of just one.
Dual access over single-view lifted Pass@1 by 8.3 to 12.1 points on OfficeQA Pro and 4.7 to 9.2 points on the APEX-Agents rubric. The resulting system, called SW-AGENT, scored 63.9% on OfficeQA Pro against a 29.3% published baseline, and 42.1% on APEX against 25.5%.
The paper doesn't break down how much of the gain comes from the hash binding versus the dual view. It also doesn't describe what a version-mismatch failure looks like in a running system, just that fixing it moves the score.
Most agent stacks I've seen treat file state as plumbing: fetch a document, hand the agent whatever text comes back, move on. StagedWorkspace's framing, that workspace state is an experimental variable you can measure and improve, generalizes past document QA. Any agent juggling multiple representations of the same file is probably leaving points on the table by not versioning them explicitly. Think a database record and a rendered view, or a diff and the file it came from. The open question is how much of that 8 to 12 point gain survives outside a benchmark built to expose exactly this failure.
Each link below shares sources, entities, or timing with this story.
Shared entity: Agent / Same source domain / Shared topic / Downstream implication
Both cover Agent; reported by the same outlet (arxiv.org); overlapping topics (agent, between, point, same).
Shared entity: Pass / Same source domain / Shared topic / Earlier coverage
Both cover Pass; reported by the same outlet (arxiv.org); overlapping topics (against, agent, artifact, same).
Salesforce released apex / Shared entity: Agent / Shared topic / Earlier coverage
Linked by a graph relationship (Salesforce released apex); both cover Agent; overlapping topics (against, agent).
Linked by a graph relationship (Salesforce released apex); both cover Agent; overlapping topics (access, agent).
Shared entity: Agent / Same source domain / Shared topic / Earlier coverage / Tension
Both cover Agent; reported by the same outlet (arxiv.org); overlapping topics (against, agent).
Shared entity: Pass / Same source domain / Shared topic / Earlier coverage / Tension
Both cover Pass; reported by the same outlet (arxiv.org); overlapping topics (against, agent).
Shared entity: StagedWorkspace / Same source / Shared topic
Both cover StagedWorkspace; cite the same source (arXiv 2608.18050); overlapping topics (agent, artifact).
Shared entity: Agent / Same source domain / Shared topic / Earlier coverage
Both cover Agent; reported by the same outlet (arxiv.org); overlapping topics (against, agent, artifact).