Agent context windows fill up by turn 18: 30 products, 8 strategies, one 30-turn ticket replayed
blaizedsouza · x · 2026-10-28
sermakarevich published a deep review of context management for AI agents:
- Covers 30 products and distills 8 distinct strategies
- Replays a single 30-turn coding ticket under each strategy for comparison
- Key finding: clearing tool results loses facts needed later (e.g. at turn 28), forcing redundant tool calls; compaction preserves them while saving 10% context
Written for engineers running agents beyond a dozen turns who have watched them forget facts read earlier — a practical reference for window-side management.
More from coding & agent
- EverOS Cloud hands-on: persistent memory lets Claude recall shop policies across new chats — Aiden_Tech_Ai · 2026-10-29
- Ziplime: new Python library for AI-driven algorithmic trading and backtesting — mdancho84 · 2026-10-29
- Vibe coding RPG: he needed a resume, so he built a resume editor first — generativist · 2026-10-29
- Code ownership in the AI era: you land it, you own it, bot approvals never count as human ones — ccerrato147 · 2026-10-29
- How Anthropic Reviews Code When Claude Writes 90% of It — ccerrato147 · 2026-10-29
- Leaked prompt makes Claude Code build and maintain its own skills automatically — Aiden_Tech_Ai · 2026-10-29