The Coding Agent Kept Context by Dropping, Not Rewriting

CliffCompaction bounds coding-agent context by truncating or dropping original material rather than rewriting it, and each pass starts from original content instead of compacting a prior compaction. The authors report cost reductions of up to 50% while maintaining or improving performance on Terminal-Bench, alongside longer KernelBench runs exceeding one million tokens. Parallel test-time scaling also shifted the reported cost-performance frontier. These results depend on the evaluated agents, benchmarks and pricing assumptions; deletion can still discard information a future step needs.