SWE-Pruner Pro: Let the Coding Agent Decide What to Keep

As coding agents handle longer and noisier tool outputs, context management becomes a practical bottleneck. SWE-Pruner Pro takes a different route from earlier pruning systems: instead of attaching an external code classifier, it asks whether the agent already encodes enough information to tell what matters.

What changes in this approach

• Pruning happens inside the agent: the model’s internal representations are reused for the pruning decision.
• Line-level keep/prune labels: a small head converts hidden states into a binary decision for each line of tool output.
• Length-aware embeddings: the method includes embeddings keyed to the number of lines in a tool output, helping the model adapt to different output sizes.

Why this is interesting

The key insight is not only that pruning is useful, but that the agent may already be doing part of the work internally while reading tool results. That means pruning can be aligned with the model’s own comprehension process rather than bolted on as a separate filter. In a coding workflow, this matters because logs, error traces, and repeated snippets often crowd out the signals the agent actually needs.

Reported results

According to the paper summary, SWE-Pruner Pro was evaluated on two open-weight backbones and four multi-turn benchmarks. It reduces prompt and completion tokens by up to 39% while maintaining task quality, with bounded inference overhead. On MiMo-V2-Flash, it also improves the SWE-Bench Verified resolve rate by 3.8% and raises Oolong long-context accuracy by 2.2 points.

Potential impact

If this design holds up broadly, future coding agents may no longer need to pass every tool response through a separate pruning stage. Instead, they can compress context more naturally as part of interpretation, which could make long debugging sessions and agentic coding loops more efficient.

In short, SWE-Pruner Pro suggests that the coding model may already know more about what to prune than we usually assume.

Source: Hugging Face Daily Papers