CORVUS Optimizes Context for LLM Coding Agents.

Large language model (LLM) coding agents typically build trajectories that accumulate reasoning steps, tool calls, and results to facilitate multi-step decision-making. However, the common append-only architecture for these trajectories often tightly links file-read actions with their observations, creating fixed snapshots in the chronological history. This design leads to inefficiencies: as files are modified by the agent or concurrently by humans, these snapshots become outdated, causing reasoning errors and forcing agents to redundantly re-read files, which further bloats the trajectory with duplicate content. To mitigate these issues, researchers propose CORVUS, a new trajectory architecture. CORVUS addresses the problem by decoupling file-read actions from their observations. Instead, it maintains a synchronized registry of relevant files and injects only their current contents at each reasoning cycle. This fundamental structural change results in significantly lighter-weight trajectories that are inherently synchronized with the actual codebase state. This approach eliminates redundant file copies and stale snapshots that typically inflate conventional trajectories. Evaluated across four LLMs on SWE-POLYBENCH_VERIFIED and SWE-BENCH PRO benchmarks, CORVUS achieved substantial reductions: 9-50% fewer average input tokens per task, 15-32% shorter final prompts, and up to 37% fewer reasoning cycles, all while preserving comparable pass rates. This demonstrates a significant improvement in efficiency for LLM coding agents.