A 2026 study ran standard token-pruning compressors on LLM agents across 17 configurations. Every one collapsed to near-zero reward, from 1.3x to 13.3x compression. The tokens a compressor drops first are the ones agents need most.
A pre-registered randomized trial compressed production agent prompts at three retention rates. Keeping half the tokens cut total cost 27.9%. Keeping a fifth raised cost 1.8%, because the model wrote longer outputs.
Engineering, research, cost, and tutorial posts in one feed.
Long-form pieces from the team building gotcontext. Compression internals, model evals, customer post-mortems. RSS feed available — see /blog/rss.xml. Unsubscribe with one click.