Veteran agent builder distrusts auto-compaction, rolls his own optimized pipeline
omarsar0 · x · 2026-11-04
Elvis (omarsar0) argues that since most agent harnesses rely on context compaction, better measurement is urgently needed — especially with longer-running tasks and multi-agent systems. He explicitly doesn't trust built-in auto-compaction and built his own, heavily optimized for his tasks and frequently assessed.
He recommends an NVIDIA paper showing that with frequent compaction, frontier models lose little to no accuracy on long-horizon and long-input tasks while keeping context short; with one action per context, Claude Opus 4.5 scores 70.2 on SWE-bench Verified.
Related event: NVIDIA Paper: Frequent Context Compaction Beats Long Contexts(3 posts)→
More from coding & agent
- LangGraph now runs everything at LangChain as team bets on deterministic agent infrastructure — LangChain · 2026-11-05
- Lovable hits 70M projects and 100K PRs as founder revisits his 2023 bet on describable software — AlexandrePesant · 2026-11-05
- Dev uses Gemini 4 + Antigravity to auto-document a week of building into a BTS video — alexanderchen · 2026-11-05
- SpecAgent: multi-session Claude Code pipeline with measured 3x context cost savings — AdOtherwise9458 · 2026-11-05
- open-duet hits 6.8k stars as self-hosted open-source alternative to Grok Bot — matchaman11 · 2026-11-05
- Solo dev open-sources a 100% local screen-monitoring agent with real-world triggers — Roy3838 · 2026-11-05