CORVUS reduces LLM coding agent tokens by up to 50% with smarter context management

A team of researchers (Zheng, OBrien, Cui, et al.) has introduced CORVUS, a novel trajectory architecture for LLM coding agents that fundamentally rethinks how context is managed during multi-step reasoning. Conventional coding agents use an append-only trajectory that captures file-read snapshots permanently in chronological order. As files are modified by the agent or concurrent human edits, these snapshots become stale, forcing agents to re-read files and append duplicate copies, bloating context windows and wasting tokens.

CORVUS solves this by decoupling file-read actions from their observations: it maintains a synchronized registry of relevant files and injects only their current contents at each reasoning cycle. This structural change produces lighter trajectories that stay synchronized with the actual codebase. On SWE-POLYBENCH_VERIFIED and SWE-BENCH PRO across four LLMs, CORVUS reduced input tokens by 9-50%, final prompt length by 15-32%, and reasoning cycles by up to 37%, while holding pass rates steady. The approach directly cuts API costs and latency for agentic coding workflows.