CliffCompaction: Cost-Efficient Compaction for Long-Horizon Coding Agents

TL;DR

CliffCompaction is a cost-efficient autocompaction method that reduces context costs by up to 50% without rephrasing or rewriting content, enabling sustained performance across long-horizon tasks. It maintains fidelity by only truncating original content and discarding prior compactions, preventing context drift. The technique improves test-time scaling efficiency and achieves state-of-the-art results on benchmark tasks with minimal cost.

Why It Matters

Long-horizon AI agents face severe context limitations, making efficient compaction critical for real-world deployment. CliffCompaction enables cost-effective, scalable reasoning over millions of tokens without performance degradation, significantly lowering operational costs for complex coding tasks.

Key Findings

• CliffCompaction reduces context-related costs by up to 50% while preserving or improving performance on Terminal-Bench and KernelBench.
• By only truncating original content and discarding prior compactions, the method avoids context drift, enabling stable learning over sessions exceeding a million tokens.
• Under test-time scaling, CliffCompaction allows Kimi K2.6 to match or exceed Opus 4.7 and GPT-5.3 Codex at lower cost, achieving 3.58× speedup on CUDA kernel generation after 400 steps.