Open-source tool-output compressor cuts coding-agent bills 22.7% without hurting solve rate
MattAlreadyHere · reddit · 2026-10-16
Hystax open-sourced a compressor for coding-agent tool output (about half of prompt bytes in their traces). Three lessons: (1) compression can raise bills — 93% of input tokens were 1/10-price cache reads, so compression must be deterministic or the cached prefix breaks and the conversation re-prices; (2) de-duplicate, don't summarize — a rewriting compressor saved 6.2 more points but lost 12.9 points of details agents reuse (paths, identifiers, versions); keep one real copy of each distinct line and mark repeats; (3) make every elision retrievable — markers like [... 412 log lines omitted [ref]; 3 ERROR] plus a fetch tool. On Terminal-Bench 2.1 (Codex CLI on GPT-5.5, 87 tasks, real bills): cost −22.7%, input tokens −25%, solve rate 82.2%→81.1% (not significant), >200K-token requests 1.9%→0.13%; biggest savings on builds/ML training (26–28%), least on short algorithmic tasks (−8%).
Related event: Hystax Open-Sources Kompressor, Cutting Coding Agent Costs by ~25%(2 posts)→
More from coding & agent
- Claude cut a 2.6-hour video export to 20 minutes — but only after being told to rebuild the server — byhewar · 2026-10-23
- 3 flight-booking rules my AI agent checks that I'd never do by hand — Efistoffeles · 2026-10-23
- OpenJet claims Qwen3.8-27B coding runs 1.9x faster than llama.cpp on one RTX 3090 — JeremyCMorgan · 2026-10-23
- Building an autonomous SEO agent in Hermes with three skills: programmatic pages, Keywords Everywhere, DataForSEO — EXM7777 · 2026-10-23
- Building a sports betting AI agent with onchain execution and human approval by default: architecture and lessons — Dry-Comfortable-2678 · 2026-10-23
- How do you audit AI agent decisions in regulated workflows? A field survey — Ok_Way_1993 · 2026-10-23