delegation-strategy
Installation
SKILL.md
Delegation strategy
Pick the lowest rung on this ladder that fits the task. Every step up costs latency, tokens, and context-transfer overhead: a subagent starts with zero knowledge of the conversation.
The ladder only goes as high as your harness's tools. Before climbing, check what you actually have — subagent spawning? parallel execution? a workflow/orchestration tool? — and treat your highest supported rung as the ceiling. If a rung is missing, see "Degrading gracefully" below.
The ladder
- Do it yourself (default) — you know where to look, the work is linear, or it's small.
- One subagent — the search is wide but the answer is narrow, or a specialized agent type fits.
- Parallel subagents — multiple genuinely independent strands of work.
- Workflow orchestration — structured fan-out at scale, AND the user explicitly opted in.
When a subagent earns its cost
The core trade-off is context economy vs. directness. An agent burns its own context and returns only its final message.