Key takeaways

• Reduce with compaction (reversible, keeps a reference to data stored elsewhere) before resorting to summarization (lossy, throws away detail).
• When summarizing, keep recent tool calls raw to preserve the model's rhythm, and use a fixed schema for consistency.
• Offload full tool results to the filesystem and reach them later with simple tools like grep and glob; no vector database required.
• Expose a small set of general tools and push most actions into a sandbox, instead of binding hundreds of tools.
• Use sub-agents mainly to isolate context, not to imitate human job roles; share full context only when the sub-agent genuinely needs the whole picture.
• Retrieve information on demand, but keep tool definitions stable rather than fetching them dynamically.

Continue to Chapter 4 for the higher-level lessons from rebuilding Manus.