How agent harness context windows work: append-only for caching and reasoning
PawelHuryn · x · 2026-11-05
Step 3 of the Mini Harness series explains the context window: models remember nothing between calls, so the harness resends everything each step — a fixed top written once per session (instructions, AGENTS.md, tool manuals, 8K tokens in practice) plus an ever-growing conversation. The key rule: never edit what was already sent, because Claude seals its reasoning to everything above it and the provider caches the prefix — editing breaks both reasoning and cache hits.
More from coding & agent
- tw93 open-sources Kami: a design system that makes AI-agent documents actually readable — sujingshen · 2026-11-05
- Power user shares the one thing he won't let Claude do: write his tweets — PrajwalTomar_ · 2026-11-05
- Old iPad reborn as a smart home wall panel, thanks to Claude — PrajwalTomar_ · 2026-11-05
- He replaced paid app subscriptions with Claude-built tools, saving $200/month — PrajwalTomar_ · 2026-11-05
- Running Claude Code on DeepSeek for chores: one job costs $0.0043 — PrajwalTomar_ · 2026-11-05
- Agent audited a 300K-star skill pack in 70 seconds, found startup hook and usage counter — PrajwalTomar_ · 2026-11-05