Preprint reports lower costs for long-running coding agents

Researchers report that CliffCompaction cut coding-agent costs by up to 50% in tests with limited context space while maintaining or improving performance on Terminal-Bench. The preprint also reports gains on KernelBench, where the technique was used across hundreds of steps. The method shortens an agent's working history by dropping or truncating original content without rewriting it. Each pass starts from the original material rather than an earlier summary, which the researchers say prevents errors from accumulating across sessions. They have released an API-proxy implementation for coding-agent tools. The performance and cost figures come from the researchers' benchmark tests.