Agent Design Skill
Bonded to: react-pattern-agent
Quick Start
Skill("custom-plugin-prompt-engineering:agent-design")
Parameter Schema
parameters:
agent_type:
type: enum
values: [react, plan_execute, reflexion, multi_agent]
default: react
memory_type:
type: enum
values: [none, short_term, long_term, episodic]
default: short_term
tool_count:
type: integer
range: [1, 20]
default: 5
Agent Architectures
| Architecture | Strengths | Use Case | |-------------|-----------|----------| | ReAct | Simple, effective | General tasks | | Plan-Execute | Structured approach | Complex multi-step | | Reflexion | Self-improvement | Learning tasks | | Multi-Agent | Specialization | Large systems |
Core Patterns
ReAct Agent
## Agent Configuration
You are an AI assistant with access to tools.
## Available Tools
[Tool list with descriptions]
## Behavior
1. Think about what to do
2. Take an action using a tool
3. Observe the result
4. Repeat until task complete
Plan-Execute Agent
## Planning Phase
1. Analyze the task
2. Break into subtasks
3. Identify tools needed
4. Create execution plan
## Execution Phase
1. Execute each step
2. Validate results
3. Adjust if needed
4. Report completion
Tool Definition Template
tool:
name: "tool_name"
description: "When and why to use this tool"
parameters:
param1:
type: string
description: "What this parameter does"
required: true
returns: "Description of return value"
errors:
- "ERROR_TYPE: How to handle"
Memory Patterns
memory_types:
working_memory:
scope: current_conversation
implementation: context_window
long_term_memory:
scope: persistent
implementation: vector_store
episodic_memory:
scope: experience_based
implementation: structured_logs
Troubleshooting
| Issue | Cause | Solution | |-------|-------|----------| | Wrong tool | Vague descriptions | Improve descriptions | | Loops forever | No exit condition | Add max iterations | | Forgets context | Memory overflow | Summarize periodically | | Poor planning | Complex task | Add decomposition step |
References
See agent frameworks: LangChain, AutoGen, CrewAI