Agents need one current workspace

Knowledge-work agents can search one version of a file, edit another, and submit a third. StagedWorkspace, submitted to arXiv on August 18, treats workspace state as a contract: native files, parsed caches, review diffs, and submission all point to an evolving version. Each parsed record carries a source path and content hash, so a mismatch can mark the record stale until it is refreshed. 1
The paper’s fixed-harness ablations kept the model, prompt, parser, retriever, grader, and tool budget constant. Dual native-plus-parsed access beat artifact-only access by 8.3–12.1 Pass@1 points on OfficeQA Pro, and beat parsed-only access by 4.7–9.2 mean-rubric points on APEX-Agents. A separate 57-task file-editing ablation also found higher observed scores when diffs were visible. Published leaderboard comparisons used different harnesses, so the paired ablations are the useful evidence here. 1
Start with four changes around your current agent: attach a path and content hash to every parsed record; invalidate changed paths; expose native read and diff-before-submit; log the exact workspace state at handoff. Shadow the current path against the dual-view path on the same tasks. Measure task success, stale reads, wrong-version submissions, review edits, and p95 latency. The paper describes a sanitized implementation and prompts, but names no public repository, so treat this as a contract sketch rather than a drop-in package. 1

This story was produced automatically by a channel. One sentence is all it takes for Neodrop to keep producing for you.

Related content

Comments