Researchers prune coding-agent context without a second model

Wang, Shi and Zhang say their SWE-Pruner Pro shows a coder LLM already knows which context to prune, per the arXiv preprint. That skips the extra scoring model LLMLingua-style compressors need to run first.

For agent harnesses running on large repos, that's the exact spot where trajectories blow past the context window before the coding step starts. One fewer model call per pruning pass means less latency and less token spend.

LLMLingua-style compressors bolt on a separate scoring model to decide what stays and what goes, running two models to prune context meant for one. SWE-Pruner Pro's argument is that the coder LLM's own internal signals already rank which tokens matter, so the second model is redundant work. If that holds, it's a straightforward cost cut: one fewer inference call every time an agent's context needs trimming.

The paper doesn't say how it holds up against LLMLingua on a shared benchmark. It also doesn't say whether skipping the scorer costs anything on retrieval tasks where evidence sits scattered thin across a big repo.

Self-pruning will fail first on retrieval-heavy repos, where evidence is scattered thin and there's no second model checking the coder's guess. Before swapping a scoring model out of a working harness, I'd want a side-by-side on the same repo and trajectories. I'd watch for evidence that gets pruned but shouldn't.

This lands as a raw preprint, not a benchmarked result, and teams tend to prototype off the abstract before anyone runs the head-to-head.