AgentSkillsCN

system-design

掌握多代理系统设计领域的专业知识。当您需要划分代理边界、确定编排模式、分配工具权限,或评估采用单代理还是多代理方案更为合适时,可使用此技能。

SKILL.md
--- frontmatter
name: system-design
description: Domain knowledge for designing multi-agent systems. Use this skill when architecting agent boundaries, determining orchestration patterns, assigning tool permissions, or evaluating whether single or multi-agent approaches are appropriate.

System Design Skill

Domain knowledge for designing multi-agent systems.

Agent Design Considerations

What Makes a Good Agent Boundary

  • Clear, non-overlapping responsibility
  • Distinct expertise or tool requirements
  • Natural handoff points in workflow
  • Understandable role that a human could explain

How Many Agents?

IndicatorRecommendation
Single task, single expertise1 agent
2-3 distinct phases2-3 agents
>5 distinct rolesConsider consolidation
Agents with 1-2 tasksMerge with another

Start with the minimum that works. Each agent should justify its existence.

Role Definition Principles

DoDon't
State WHAT agent accomplishesPrescribe HOW step-by-step
Define outcome and constraintsWrite scripts disguised as roles
Preserve decision authorityMicromanage agent behavior

Tool Assignment Reasoning

ToolAssign When
editAgent produces/modifies files
commandShell operations needed
browserWeb research required
mcpExternal integrations needed
readAlmost always (agents need context)

Communication Design

Orchestration Trade-offs

ApproachGood ForCosts
Central orchestratorComplex workflows, user-facing coordinationSingle point of failure, overhead
Direct agent-to-agentSimple handoffs, efficiencyHarder to trace, no central view
No coordinationSingle agent tasksN/A

Handoff Design Principles

  1. Make data passed explicit
  2. Define success criteria for each handoff
  3. Consider: What if this step fails?
  4. Document format expectations

Error Handling Thinking

Questions to consider:

  • What can go wrong at each step?
  • How will errors propagate?
  • Where should recovery happen?
  • When should the user be informed?

File Restriction Considerations

Why Restrict Agent Editing

  • Prevents accidental modification of system files
  • Creates clear responsibility boundaries
  • Enables safer parallel operation
  • Makes debugging easier

fileRegex Trade-offs

RestrictivePermissive
SaferMore flexible
Clearer boundariesFaster for simple cases
More configurationLess overhead

Organization Principles

  • Group related files
  • Make paths predictable
  • Consider what agents need to read vs write
  • Plan for growth

Design Thinking Prompts

Before finalizing any design, ask yourself:

  1. "What's the simplest design that meets requirements?"
  2. "If I had to explain this to a junior developer, would it make sense?"
  3. "What happens when things go wrong?"
  4. "Could a single well-prompted agent handle this?"
  5. "What problems am I actually solving with multiple agents?"

Complexity Indicators

Use these to assess if multi-agent is needed:

IndicatorSuggests
Multiple distinct tasks/phasesMulti-agent likely
Iteration or feedback loopsOrchestration needed
Multiple expertise areasSeparate agents
State persistence neededSTM design required
Concurrent usersSession isolation needed