Agent Architecture Deep Research & MVP Benchmarks

Release: v0.0.1 | Updated: 2026-04-26 13:00 (UTC+8) | Maintainer: mambo & ca

---

🚀 Overview

This module contains deep-dive research into the core mechanisms of top-tier open-source Agent frameworks (AgentScope, LangChain, etc.), combined with pseudo-code implementations and MVP (Minimum Viable Product) benchmarks to validate architectural decisions.

Target: To dissect, understand, and reproduce the internal logic of modern Agent systems, from communication protocols to memory management.

📦 Module Manifest

Current directory structure and core research subjects:

Module	Focus Area	Validation Artifact
channel/	Gateway, Heartbeat & Debounce	mvp_qwenpaw_channel.py
communication/	Multi-Agent Protocols (Go/Python)	benchmark_communication.go
evaluate/	LLM-as-Judge & Cost Metrics	mvp_evaluation.py
mcp_protocol/	MCP Client & Tool Discovery	mvp_mcp.py
memory/	Dream Optimization & Vector Retrieval	mvp_memory_dream.py
pipeline/	Workflow Orchestration (Fan-out/Seq)	mvp_pipeline.py
planning/	Explicit Plans & Memory Compression	mvp_planning.py
security/	Async Approval & Circuit Breakers	mvp_approval_system.py
skills/	Skill Pool & Env Isolation	mvp_skill_loader.py
subagent/	Sub-Agent Lifecycle & Sandbox	mvp_subagent_lifecycle.py
trace/	OpenTelemetry & Observability	mvp_tracing.py

(Note: mul-agent/ contains historical communication baselines and sub-repo research.)

💡 Key Findings

1. Priority Routing (Channel): Implemented a 3D priority routing system (Source/Context/Urgency) to prevent queue blocking.
2. Async Approval (Security): Modeled on asyncio.Future to allow non-blocking user interaction with strict timeout circuit breakers.
3. Dream Optimization (Memory): Automated background summarization turns chaotic logs into structured, persistent knowledge.
4. Protocol Unification (MCP): Standardized tool exposure via JSON Schema extraction, decoupling Agent logic from Tool implementation.

🛠️ Local Dev Guide

• Python: uv run <module>/mvp_*.py
• Go: go run <module>/benchmark_*.go

---

📖 项目概述 (中文)

本项目针对主流开源 Agent 框架的核心机制进行源码级调研，并通过伪代码复现与MVP 基准测试验证架构设计的可行性。

🔧 模块清单

• channel/: 网关心跳与消息防抖
• communication/: 多智能体通信协议与并发基准
• evaluate/: LLM 裁判评估体系与成本监控
• mcp_protocol/: MCP 协议接入与工具动态发现
• memory/: 记忆管理与做梦优化机制
• pipeline/: 串行/并行工作流编排
• planning/: 任务规划与结构化压缩
• security/: 异步审批与超时熔断
• skills/: 技能池加载与环境隔离
• subagent/: 子代理生命周期管理与沙箱
• trace/: 全链路追踪与可观测性

🌟 核心发现

• 采用了基于优先级的三维路由防止消息堆积。
• 使用 asyncio.Future 实现无阻塞的安全审批机制。
• 通过后台“做梦”整理碎片化日志为结构化知识。
• 借助 MCP 协议统一工具暴露标准，实现逻辑解耦。