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.