AI Agent Framework Analysis¶
Deep Code Audit of 12 Open-Source Agent Frameworks
After auditing ~400KB of source code from 12 mainstream agent frameworks, one counter-intuitive fact emerges:
The core Agent Loop logic is highly similar across all frameworks. The real differences lie in "Capability Differentiation" — 35+ engineering features like state management, security controls, MCP support, and context compaction.
Quick Navigation¶
-
Blog
Core insights from the audit — academic vs. industrial capability gap explained
-
Comparison Matrix
Detailed feature comparison of all 12 frameworks across 35+ engineering capabilities
-
Deep Dive Reports
Source-code-level analysis for each framework, with file paths and line numbers
Key Insights
- Layer 1 (Core Loop): No differentiation — all 12 frameworks use the same ReAct-style pattern
- Layer 2 (Engineering Capabilities): Where real competition happens — 35+ features
- Layer 3 (Protocol Support): MCP becoming the standard (11/12 frameworks)
- Layer 4 (Interaction Mode): Just the surface layer
Frameworks Analyzed¶
| Framework | Language | Core Positioning | Stars |
|---|---|---|---|
| OpenAI Agents SDK | Python | Production SDK | 19.2K |
| Claude Agent SDK | Python | Claude Integration | 5K |
| Codex CLI | Rust | Enterprise Security | 62.2K |
| OpenCode | TypeScript | Open Source CLI | 112K |
| Kimi CLI | Python/TS | IDE Integration | 5.9K |
| Gemini CLI | TypeScript | Google Ecosystem | 95.9K |
| Qwen Code | TypeScript | Qwen Ecosystem | 18.1K |
| SWE-agent | Python | Research | 18.4K |
| OpenManus | Python | Quick Experiment | — |
| Aider | Python | Git-Native Coding | 41K |
| Goose | Rust | MCP-Native Framework | 31.4K |
| OpenHands | Python | Full Platform | 68.3K |
Research assisted by OpenCode. MIT License.