| 项目 | 主模式 | 循环方式 | 状态管理 | 并行性 | 可组合性 |
|---|---|---|---|---|---|
| Aider | Strategy ReAct | 单步迭代 | Git + Chat History | ❌ | 中 (edit-format) |
| SWE-agent | ReAct | Shell 交互循环 | Shell State | ❌ | 低 |
| OpenHands | ReAct | Event Loop | State + Condenser | ✅ (Sandbox) | 高 (Agent Hub) |
| Codex CLI DEEP | ReAct SQ/EQ | SQ/EQ Protocol Loop | Thread + Turn State | ✅ (Sandbox+Guardian) | 中 (Dynamic Tools+MCP) |
| AutoGPT | Plan-Execute | Think→Act→Observe | Memory + State | ⚠️ | 中 (Blocks) |
| MetaGPT | Actor | Role._think()→_act() | Environment + Message | ⚠️ | 高 (SOP) |
| Agent Zero | ReAct | Step Loop | Hierarchical Memory | ✅ (Sub-agents) | 中 |
| AutoGen | Actor | Message Passing | Conversation + Context Vars | ✅ | 极高 |
| CrewAI | Actor | Process-driven | Task State + Memory | ⚠️ | 高 (Flows) |
| LangGraph | State Machine | Graph Traversal | TypedDict State + Checkpoint | ✅ (Fan-out) | 极高 |
| PydanticAI | State Machine | Graph Nodes | GraphAgentState | ❌ | 高 (Type-safe) |
| Agno | ReAct | Simple run() | Session + Memory | ✅ (Multi-agent) | 高 (Toolkit) |
| Google ADK | ReAct | Runner.run_async() | Session + State | ✅ (A2A) | 高 (Tools) |
| Smolagents | ReAct | Step-based | History | ⚠️ (Managed) | 中 |
| CAMEL | Actor | Turn-taking | Chat Memory + Knowledge | ✅ | 高 |
| Browser Use | ReAct | See→Think→Act | Browser State | ❌ | 低 |
| GPT Researcher | Pipeline | Research→Report | Vector Store | ✅ (Multi-agent) | 中 |
| Goose DEEP | ReAct ACP | Streaming BoxStream<AgentEvent> | Session + Auto-Compaction | ✅ (Subagent) | 高 (ACP+Extensions) |
最成功的 Agent 往往不只用一种模式:
结论:核心是 ReAct 循环,外层用更复杂的编排模式包装。