Subagents, skills and code mode

A single agent with one long context degrades as the work grows: every document, tool definition and intermediate result competes for the same window, costs tokens on every step, and distracts the model. Three patterns decompose the work. Subagents give a sub-task its own clean context and a narrow brief, and return a short structured result, so the orchestrator never sees the raw material. Skills package instructions, scripts and reference files so the agent loads only a one-line description up front and the full procedure only when the task needs it. Code mode presents tools as a code API: the model writes a short program that calls tools and filters results in a sandbox, so thousands of rows never enter its context. Each pattern trades something: subagents cost more total tokens and add coordination, skills need curation and review, code mode needs a real sandbox. Use them when the work is genuinely separable or large; keep a single agent when it is small, tightly coupled or hard to debug.

Key takeaways:

• Subagents reduce peak context, not total cost. Expect more total tokens and more calls.
• A subagent's brief is a contract: objective, inputs, output schema, tools, boundaries.
• Skills are progressive disclosure for procedures: metadata always, body on activation, files on demand.
• Code mode keeps intermediate data out of the model's context, and makes sandboxing non-negotiable.
• Decompose for isolation, parallelism or scale; never for elegance.