AgentSkillsCN

deepagents-harness

对于那些受益于任务委派、并行研究或结构化规划的多任务场景,此技能可助您事半功倍。它引导主代理善用任务工具、子代理,以及write_todos功能,实现高效协作与任务统筹。

SKILL.md
--- frontmatter
name: deepagents-harness
description: Use this skill when the user asks about deepagents harness capabilities, tool names, task delegation, filesystem access, or sandbox execution. It provides a concise operational cheat sheet.

DeepAgents Harness Skill

Core Capabilities

  • Task delegation: Use task to spawn subagents for isolated work.
  • Planning: Use write_todos for multi-step tasks and update statuses.
  • Filesystem tools: ls, read_file, write_file, edit_file, glob, grep.
  • Human-in-the-loop: interrupt_on gates sensitive tools.
  • Summarization: History auto-compresses at high token usage.

Operational Guidance

  • Delegate early when work can be parallelized (research + analysis + drafting).
  • Keep tasks small, with explicit output format.
  • Use file tools for structured artifacts and breadcrumbs.
  • Prefer short, actionable plans with 3-6 steps.

Common Patterns

  • Research: plan → task subagents → synthesize with citations.
  • Automation: plan → run sandbox workflow → verify → summarize.
  • Comms: draft → confirm → send.