Agent Architecture Patterns

AI Agent architecture patterns library with 10 patterns for single and multi-agent systems

Security Scan

The skill's files and runtime instructions match its stated purpose (an architecture patterns library with examples); nothing requests unrelated credentials or external installs, though some example code contains patterns that are risky to run with real data (eval, logging secrets).

Purpose & Capability

The name/description (agent architecture patterns) align with the included files: pattern docs, example apps, agent implementations and tests. The files present are expected for a patterns + examples library.

Instruction Scope

SKILL.md only instructs reading the patterns, asking the agent questions, and running bundled examples/tests. It does not request secrets or external endpoints. Note: several example files intentionally demonstrate insecure code (e.g., printing a user object containing a password and using eval for expression evaluation); the runtime instructions encourage running examples, so users should not run examples on sensitive data or without review.

Install Mechanism

No install spec is provided (instruction-only). The package contains code but there is no download/install step that fetches remote artifacts — lowest install risk.

Credentials

The skill declares no required environment variables, no primary credential, and no config paths. Nothing requests disproportionate credentials or access.

Persistence & Privilege

always is false and model invocation is allowed (platform default). The skill does not request persistent system-wide privileges or try to modify other skills' configuration.

Assessment

This package appears to be a coherent patterns + examples library. Before running anything: 1) review example code for risky constructs — e.g., examples/react-example.js uses eval(expression) which can execute arbitrary code, and examples/manager-worker-full-example.js logs a user object (including password) — do not run examples on real data. 2) Run examples/tests in a sandbox or disposable environment. 3) If you plan to reuse code, replace eval with a safe expression parser and remove any sample logging of secrets. 4) The package.json repository URL is a placeholder; if provenance matters, verify the source/maintainer before trusting this code in production.

This skill provides a comprehensive library of AI Agent architecture patterns to help developers:

• Design single-agent architectures (ReAct, Reflection, Self-Critique, Plan-and-Solve, Tree of Thoughts)
• Design multi-agent collaboration systems (Manager-Worker, Peer-to-Peer, Hierarchical, Market-Based, Pipeline)
• Apply system design principles (separation of concerns, fault tolerance, scalability)
• Implement best practices based on OpenClaw

---

Patterns

Single-Agent Patterns (5)

1. ReAct - Reasoning + Acting alternation
2. Reflection - Self-reflection and iterative improvement
3. Self-Critique - Self-criticism and error correction
4. Plan-and-Solve - Plan first, then execute
5. Tree of Thoughts - Multi-path exploration

Multi-Agent Patterns (5)

1. Manager-Worker - 1 manager coordinates multiple workers
2. Peer-to-Peer - Equal agents collaborate
3. Hierarchical - Multi-level management structure
4. Market-Based - Task bidding and allocation
5. Pipeline - Sequential multi-stage processing

---

Usage

Option 1: Consult AI-Agent

Ask questions like:

• "Design a multi-agent code review system"
• "How to implement ReAct pattern?"
• "Which agent collaboration pattern should I use?"

Option 2: Reference Documentation

Browse patterns/ directory for detailed pattern docs.

Option 3: Use Code Examples

Run example code from examples/ directory.

---

Examples

ReAct Pattern Example

const agent = new ReActAgent({
  tools: [search, calculate],
  maxSteps: 10
});

const answer = await agent.execute("What's the temperature in Beijing today?");

Manager-Worker Pattern Example

const workers = [
  new WorkerAgent('worker-1', ['javascript'], { codeReview: true }),
  new WorkerAgent('worker-2', ['python'], { dataAnalysis: true })
];

const manager = new ManagerAgent(workers);
const result = await manager.coordinate("Review this codebase");

---

Installation

clawhub install agent-architecture-patterns

---

Testing

npm test
# Runs 30 test cases for ReAct and Manager-Worker implementations

---

License

MIT

---

Author

AI-Agent