When an LLM-based coding agent works on a codebase, its behavior resembles that of a novice programmer: it reads a file, scribbles it into the history, and then forgets that the information has changed. The conventional append-only trajectory model pastes every file read as a snapshot into the chronology. If the agent itself or a human collaborator later edits that file, the snapshot becomes stale. Confused, the agent re-reads the same file multiple times, bloating the history with redundant copies and wasting precious tokens. CORVUS, an architecture introduced by a research team, addresses the problem at its root: instead of accumulating reads, it maintains a synchronized registry of relevant files and injects only the current content at each reasoning step.

The mechanism is simple yet radical. Rather than treating each file-read as a historical event to be appended, CORVUS decouples it from the observation: the agent knows it read a file, but the actual content is fetched from the registry at the next reasoning cycle. This ensures the trajectory stays aligned with the real codebase state, without the redundancy of multiple snapshots. Tests on SWE-POLYBENCH_VERIFIED and SWE-BENCH PRO with four LLMs showed substantial savings: 9-50% fewer input tokens per task, 15-32% shorter final prompts, and up to 37% fewer reasoning cycles, all while maintaining comparable pass rates.

For those evaluating on-premise deployment of coding agents, these numbers carry significant weight. Every token saved reduces the computational load during inference: less VRAM occupied by the context, lower latency, and above all, lower operational costs. In a self-hosted scenario where capital-intensive GPUs must deliver maximum throughput, halving tokens means handling more concurrent tasks or extending usable context windows without saturating memory. CORVUS requires no model changes or additional hardware: it’s a purely architectural intervention on the trajectory, applicable to any running LLM.

There is also a data sovereignty angle. In traditional systems, the multiplication of snapshots not only wastes tokens but spreads copies of the codebase across the pipeline. In a regulated setting (GDPR, proprietary data), limiting this proliferation reduces the exposure surface and simplifies audits. By maintaining a single synchronized registry, CORVUS curbs the dispersion of sensitive information – a side benefit for air-gapped environments or those with strict data residency requirements.

The architecture signals a mindset shift: it’s not about making models more powerful, but about making them more efficient downstream of inference. This is the same principle that drove quantization and pruning: cut waste without sacrificing quality. CORVUS shows that trajectory management is fertile ground for optimization, with immediate payoffs for local pipelines. While cloud vendors push for pricier tokens, solutions like this make on-premise inference increasingly attractive, where cost control is direct and every optimization translates into better operating margins.