Compaction Is Good Enough Now: Agent Context Hygiene Tips From the Field
garrytan · x · 2026-11-08
dexhorthy argues that with compaction much improved, explicitly clearing or externalizing context for model focus is often unnecessary—still useful for human collaboration or agent handoffs. Practical numbers: for 1M-context frontier models he autocompacts to 400k and manually compacts between 200k–400k depending on model/harness. He also backs 3-tier subagent architectures where a top-level orchestrator launches subagents that can launch their own, noting little reason to nest beyond three levels today.
More from coding & agent
- springrod critiques Snowflake's context layer: SQL-only, single account, no recursion — springrod · 2026-11-08
- Cross-Validate With Subagents on Different Models via Vercel AI Gateway — steren · 2026-11-08
- Blog series explains how AI agents know when a task is actually done — PerfectAd4814 · 2026-11-08
- Twombly launches: a shared canvas for collaborative design with AI agents — floguo · 2026-11-08
- Building an art historian agent: which scraping API survives long-term, Tavily or Apify? — raedyohed · 2026-11-08
- GraphRAG vs RAPTOR vs HippoRAG: how to pick a RAG architecture for AI agents — MaryamMiradi · 2026-11-08