Teknium explains why "tool-call pruning" compaction keeps getting rejected from Hermes Agent
Teknium · x · 2026-09-20
Teknium (co-founder of Nous Research) broke down how a so-called "jev compaction" scheme works and why it fails:
- At 500K context, it removes 250K tokens of tool calls, compacting back to 250K — not the 50K a summary-based approach achieves.
- Each subsequent compaction can remove less and less (125K, then 75K...) as tool-call tokens deplete, until nothing is left to prune.
- So it cannot keep a session going forever and costs more than summarizing the full context once into a small block.
He noted this approach has been PR'ed to Hermes Agent many times, programmatically, and rejected every time for these reasons.
More from coding & agent
- Agentic Engineer: Scaling Trust in Your Agents Is the #1 Goal — brandon_galang · 2026-09-20
- Bespoke Labs open-sources Nimble, an open Jev rival for fast typed decisions — RexDouglass · 2026-09-20
- 21M tokens for just $0.76: dev shares eye-popping AI coding bill — aronchick · 2026-09-20
- Compound Engineering adds "bakeoff": agents must build competing approaches before committing — danshipper · 2026-09-20
- Dev Builds Mnemos Cognitive Architecture and a "Retirement Home" for Deprecated AI Models — RileyRalmuto · 2026-09-20
- ChatGPT agent inspected repo, filed 3 GitHub issues and assigned them all to Copilot — dfinke · 2026-09-20