Tips: Long Sessions Waste Tokens and Degrade Model Performance
hudsonhateno · reddit · 2026-09-02
A Reddit user explains two main issues with long-running LLM sessions: token consumption grows linearly with context processing, and model output quality degrades significantly once context usage exceeds 40%. The post lists context windows and max output limits for various Codex models and recommends "handing off" sessions—summarizing the current state and starting a new one—when capacity reaches 40%-50%. The author also shares two custom skills: $context-capacity to calculate current context load and cumulative usage, and $handoff to assist in the session handoff process.
More from coding & agent
- LunagraphHQ Launches: Visual Editor for React — floguo · 2026-09-02
- Rippling Launches AI Agent to Automatically Resolve Tedious IT Issues — andreisavu · 2026-09-02
- Nested traces are outdated; observability tools should focus on chat views and failure surfacing — _ScottCondron · 2026-09-02
- Backend Consensus Shift: Use Effect Instead of Plain TypeScript — mattpocockuk · 2026-09-02
- Entrepreneur shares Grok Bot workflow: automates computer, tests apps, learns workflows in one evening — PrajwalTomar_ · 2026-09-02
- Paper coding forces attention to detail over autocomplete and LLM reliance — TivadarDanka · 2026-09-02