AINEWS
An arXiv preprint called CORVUS reports 9–50% less text fed to the model per coding task, by separating the act of opening a file from the stale copy the agent has been carrying in its working memory.
AI coding agents like Cursor, Claude Code, and Devin-style tools re-read files they already have in context, and each re-read gets permanently appended to the agent's running history: the trajectory of reads, tool calls, and reasoning steps the model accumulates across a task. A new arXiv preprint argues that append-only coupling is the reason these agents burn extra compute on redundant work, and proposes a fix the authors say trims input tokens by 9–50% per task.
The paper, CORVUS: Context Optimization and Reduction Via Underlying Synchronization for LLM Coding Agents, decouples file reads from their observations by maintaining a synchronized registry of relevant files and injecting only their current contents at each reasoning cycle. Conventional agents fold each file snapshot into the chronological trajectory, where it stays even after the file changes on disk. So a Read on line 10 of a task can still be holding a stale copy of auth.py at the moment the agent edits that file on line 200, and the only way the model finds out is by re-reading. CORVUS replaces that snapshot with a live pointer, so the trajectory stays in sync with the codebase by construction.
On SWE-PolyBench Verified and SWE-Bench Pro, evaluated across four LLMs the paper does not name, the authors report 9–50% fewer input tokens per task, 15–32% shorter final prompts, and up to 37% fewer reasoning cycles. The catch is explicit in the paper: pass rates were "comparable" rather than improved. CORVUS is a cost lever, not a capability claim.
The timing is what makes the paper legible rather than academic. Cursor, Claude Code, Devin-style tools, and a long tail of open-source coding agents are now hitting token-cost ceilings at scale, and a trajectory-level diagnosis is more useful than a per-model patch. The current workaround is bigger context windows, longer-token frontier models, and aggressive prompt caching, all of which raise the bill at a different layer. If a slice of that cost comes from redundant re-reads, a slice the paper argues is common to current architectures, fixing the trajectory compounds across every model the agent runs, including the smaller and cheaper ones the industry is migrating toward.
The paper is an arXiv preprint, not peer-reviewed. The four evaluated LLMs are not named in the abstract, so the 9–50% range is an aggregate across models and benchmarks rather than a per-model number, and the per-benchmark breakdown lives in the full PDF. There is no production deployment, no independent replication, and the headline "comparable pass rates" is the paper's own framing, not a soft launch.
The wire version of this story will say "new paper trims AI coding-agent input by up to 50%" and stop. CORVUS is one move inside a larger push for cheaper, leaner coding agents, alongside context-window hacks, retrieval-augmented agents, and shorter tool-call prompts. Three follow-ups will tell readers whether the result holds: which four models the paper actually used, the per-benchmark breakdown, and whether any independent lab reproduces the trajectory compression on production coding-agent logs.