Scaling Long-Horizon Coding Agents

Test-time scaling works well for short, comparable outputs, but long-horizon coding agents leave behind long rollouts of actions, observations, errors, and partial progress. In this setting, what matters is not more attempts, but a representation that lets you select and reuse prior experience.

Each rollout is compressed into a structured summary that keeps only the key hypotheses, progress, and failure modes while stripping out noisy execution details. Based on this representation, two inference scaling approaches become possible.

• Parallel scaling: Recursive Tournament Voting (RTV) recursively repeats small-group comparisons to progressively narrow down candidates.
• Sequential scaling: Parallel-Distill-Refine (PDR) generates new rollouts conditioned on summaries distilled from previous attempts.

This approach consistently improved overall coding agent performance on SWE-Bench Verified and Terminal-Bench v2.0. For example, Claude-4.5-Opus rose from 70.9% to 77.6% on SWE-Bench Verified using mini-SWE-agent, and from 46.9% to 59.1% on Terminal-Bench v2.0 using Terminus 1.

Ultimately, test-time scaling for long-horizon agents comes down not to stacking more rollouts, but to how well you represent, select, and reuse experience.