Cut Agent Token Costs 95%: Hand Off at 50-60% Context Instead of Waiting for Compaction
tomcrawshaw01 · x · 2026-10-29
Tom Crawshaw hammers a key agent practice: never wait for context compaction before starting a fresh session.
- At 50%-60% context usage, create a handoff file and start a new session carrying over only what matters
- Otherwise the model gets 'dumber' and forgets information mid-chat
His article 'The 7 Changes That Make Hermes Agent Up to 95% Cheaper To Run' shows a free audit plugin that cut per-run overhead from 40,000 tokens to 19,200 before changing a single model, with the full step-by-step setup and a Token Optimizer walkthrough on YouTube.
More from coding & agent
- Stack Overflow launches Scopes: permission-governed knowledge spaces for AI agents — pchandrasekar · 2026-10-29
- Running NVIDIA's Open-H3-Regenerate-2K on a 4090: sharper but rewrites the actor's performance — friedlc · 2026-10-29
- mcp-coatcheck: A Proxy That Runs Slow MCP Tool Calls in the Background to Avoid Timeouts — Awacer · 2026-10-29
- Vercel launches CI that skips irrelevant tasks automatically, 2-6x faster — cramforce · 2026-10-29
- AlsoAsked MCP Server gives Claude access to Google's People Also Ask data — modelcontextprotocol · 2026-10-29
- Lobsters MCP Server pipes lobste.rs stories and discussions into MCP clients — modelcontextprotocol · 2026-10-29