A clearer view of AI.
THE INDEPENDENT AI.

WORTH TESTING · 24 OCTOBER 2026

Anthropic's Lydia Hallie on cutting Claude Code costs (the post's '65 agents in a loop' claim is false)

Watched in full (27 minutes). This is a re-upload of an official Claude Code workshop by Lydia Hallie from Anthropic's Claude Code team; the post's claim about an engineer running 65+ Claude agents in a loop is clickbait, because the talk is really about token costs, prompt caching and keeping context lean. Verdict: excellent, no-hype advice from the source; intermediate to advanced and aimed at people coding with Claude Code, though the cost lessons apply to anyone paying per token. It has ten things to test.

Try this

  1. Match the model to the task instead of always using the biggest one. 07:44
  2. @-mention files so Claude doesn't waste a step searching for them. 09:08
  3. Run /context in a fresh session to see what's bloating every request. 09:58
  4. Keep CLAUDE.md short and move specific procedures into skills. 10:44
  5. Use the /claude-api prompt-audit skill to remove outdated instructions. 11:31
  6. Switch off MCP servers you're not using, or use command-line tools like gh instead. 11:51
  7. Make test output quiet, e.g. npx vitest run with --reporter=dot. 12:17
  8. If the model fails, switch up a model; if it's lazy or stops early, raise /effort. 14:33
  9. Run /compact before changing model or stepping away. 15:42
  10. Send noisy research to subagents so your main session stays clean. 20:23

Timestamps point to the moment in the source video.

Cost

Free to watch; the advice is about lowering your own token bill.

Source

Hype or Fact?

Hype 'An engineer runs 65+ Claude agents in a loop'

Related AI tools

How we knowSteps to try, not a claim that we independently tested every result. Follow the linked sources and check current prices. Our methods & sources ↗