Context fails evolve each model generation: poisoning gave way to inertia and harness design
dbreunig · x · 2026-10-22
Ahead of AGNTCon, dbreunig shares slides on context fails and fixes, tracing a generational cycle:
- Longer contexts (GPT-4o → Gemini 1.5's 1M → Llama 4's 10M) introduced context poisoning, rot, and confusion
- Harnesses emerged to manage information; long horizons (METR: GPT-5 3.4h → Gemini 3.1 6.4h → Opus 4.6 12h) then surfaced context inertia and fighting the weights
- Now harnesses must learn when to 'forget' and manage tool loadouts
- He asks what problems the next generation will create
More from coding & agent
- Open-source huashu-motion turns Claude Code into a launch-film studio, 1.3k stars — AlchainHust · 2026-10-22
- Thinking Machines Ships StableReinforce, a Loss That Keeps Off-Policy RL From Collapsing — LiTianleli · 2026-10-22
- phoneuse launches MCP/API giving AI agents their own cloud Android and iPhone devices — Rasmic · 2026-10-22
- 15-person Keenable beats $350M-funded Exa atop Vals web search index at 49.1% — EXM7777 · 2026-10-22
- Rork's $1800/month Max tier offers 10,005 credits for tokenmaxxing Opus — rudrank · 2026-10-22
- AI Agents Spent $47.35 Overnight From a Single Prompt With No Numbers — aakashgupta · 2026-10-22