Skip to slide
29 / 29

CH 04 · Practical Lessons from Building Manus · 08/08

Read as one page

Key takeaways

  • Keep the toolset small with a hierarchical action space: a few atomic tools, sandbox utilities reached through Bash, and code for multi-step logic.
  • Treat sub-agents as tools (define a goal, tools, and output schema) rather than as a chatty org chart.
  • Do not fine-tune or train your own policy prematurely; the Bitter Lesson says a better general model will likely make that work obsolete.
  • Test your agent across model strengths; if a stronger model does not help, your harness is the bottleneck.
  • Prefer binary, computable verification (the Intern Test) over subjective scoring.
  • Expect to rebuild often, and treat growing harness complexity as a warning sign of over-engineering.
  • Get out of the model's way: minimal effective context, plus sensible safety guards on shell and browser access.

This completes the Context Engineering chapters. For the multi-agent ideas referenced here (sharing context, sub-agents, orchestration), see the multi-agent folder.

Finished this idea?

Your learning progress stays in this browser.