Agent Designer Skill
Design robust multi-agent systems: define agent roles, choose an architecture pattern, design tool interfaces, add safety guardrails, and evaluate performance.
Workflow
- Define roles -- Identify distinct agent responsibilities. Each agent should have a single clear purpose.
- Choose architecture -- Select the orchestration pattern that fits the problem. See architecture-patterns.md.
- Design tools -- Define the tool schemas agents will use. See tool-design.md.
- Add guardrails -- Layer in validation, human-in-the-loop gates, and failure handling. See guardrails.md.
- Evaluate -- Measure task completion, quality, cost, and latency. Iterate on weak points.
Agent Role Definition
Each agent should be specified with:
Agent: code-reviewer
Purpose: Review code changes for bugs, security issues, and style violations
Inputs: diff (string), file_context (string[]), review_guidelines (string)
Outputs: findings (Finding[]), verdict (PASS | FAIL), summary (string)
Tools: read_file, grep_codebase, run_linter
Constraints: read-only access, no code modification, max 5 tool calls per file
Keep agent scope narrow. An agent that "does everything" is just a prompt with no structure. If an agent needs more than 5-7 tools, it is probably two agents.
Architecture Selection Guide
| Agents |
Complexity |
Pattern |
Example |
| 1 |
Low |
Single Agent |
Code reviewer, summarizer |
| 2-3 |
Medium |
Pipeline |
Research -> Draft -> Edit |
| 3-5 |
Medium |
Supervisor |
Manager delegates to workers |
| 5+ |
High |
Hierarchical |
Multi-level delegation |
| Dynamic |
High |
Swarm |
Agents spawn/recruit as needed |
See architecture-patterns.md for detailed trade-offs.
Output Format
System: PR Review Pipeline
Architecture: Pipeline (3 stages)
Agents: code-analyzer -> security-scanner -> summary-writer
Tools: 8 total (3 shared, 5 agent-specific)
Guardrails: input validation, output filtering, 30s timeout per agent
Estimated cost: ~$0.05 per PR (avg 500 lines)
What You Get
- An agent architecture document specifying each agent's role, inputs, outputs, tools, and constraints.
- An orchestration pattern recommendation (pipeline, supervisor, hierarchical, or swarm) with rationale.
- Tool schemas, guardrail definitions, and estimated per-invocation cost for the complete system.
Sub-files
| File |
Content |
| architecture-patterns.md |
Orchestration patterns, communication |
| tool-design.md |
Schema design, error handling, idempotency |
| guardrails.md |
Safety, memory, failure handling, eval |
1---2name: agent-designer3description: Multi-agent system design with orchestration patterns, tool schemas, and guardrails. TRIGGER when: user asks to design an agent system, choose agent architecture, define tool schemas, add guardrails, or evaluate agent performance; user runs /agent-design or /agent-architect. DO NOT TRIGGER when: using a specific agent framework (use framework docs), writing prompts for single-turn LLM calls, general API integration without agent orchestration.4---56# Agent Designer Skill78Design robust multi-agent systems: define agent roles, choose an architecture pattern, design tool interfaces, add safety guardrails, and evaluate performance.910## Workflow11121. **Define roles** -- Identify distinct agent responsibilities. Each agent should have a single clear purpose.132. **Choose architecture** -- Select the orchestration pattern that fits the problem. See [architecture-patterns.md](architecture-patterns.md).143. **Design tools** -- Define the tool schemas agents will use. See [tool-design.md](tool-design.md).154. **Add guardrails** -- Layer in validation, human-in-the-loop gates, and failure handling. See [guardrails.md](guardrails.md).165. **Evaluate** -- Measure task completion, quality, cost, and latency. Iterate on weak points.1718## Agent Role Definition1920Each agent should be specified with:2122```23Agent: code-reviewer24Purpose: Review code changes for bugs, security issues, and style violations25Inputs: diff (string), file_context (string[]), review_guidelines (string)26Outputs: findings (Finding[]), verdict (PASS | FAIL), summary (string)27Tools: read_file, grep_codebase, run_linter28Constraints: read-only access, no code modification, max 5 tool calls per file29```3031Keep agent scope narrow. An agent that "does everything" is just a prompt with no structure. If an agent needs more than 5-7 tools, it is probably two agents.3233## Architecture Selection Guide3435| Agents | Complexity | Pattern | Example |36|--------|------------|-----------------|--------------------------------|37| 1 | Low | Single Agent | Code reviewer, summarizer |38| 2-3 | Medium | Pipeline | Research -> Draft -> Edit |39| 3-5 | Medium | Supervisor | Manager delegates to workers |40| 5+ | High | Hierarchical | Multi-level delegation |41| Dynamic| High | Swarm | Agents spawn/recruit as needed |4243See [architecture-patterns.md](architecture-patterns.md) for detailed trade-offs.4445## Output Format4647```48System: PR Review Pipeline49Architecture: Pipeline (3 stages)50Agents: code-analyzer -> security-scanner -> summary-writer51Tools: 8 total (3 shared, 5 agent-specific)52Guardrails: input validation, output filtering, 30s timeout per agent53Estimated cost: ~$0.05 per PR (avg 500 lines)54```5556## What You Get5758- An agent architecture document specifying each agent's role, inputs, outputs, tools, and constraints.59- An orchestration pattern recommendation (pipeline, supervisor, hierarchical, or swarm) with rationale.60- Tool schemas, guardrail definitions, and estimated per-invocation cost for the complete system.6162## Sub-files6364| File | Content |65|--------------------------------------------------------|--------------------------------------------|66| [architecture-patterns.md](architecture-patterns.md) | Orchestration patterns, communication |67| [tool-design.md](tool-design.md) | Schema design, error handling, idempotency |68| [guardrails.md](guardrails.md) | Safety, memory, failure handling, eval |
DROOdotFOO