What happens when enterprise requirements hit Strands, LangGraph, and CrewAI - 45 runs measured

说明与用途

三种代理框架在需要审计和人工批准的新闻摘要任务中进行了测试。Strands 有时会输出空结果，LangGraph 提供完整审计，CrewAI 简单但在拒绝时可能循环。

问题

Enterprise AI systems must satisfy audit trails and human approval gates, but agent frameworks differ in how they enforce these constraints.

结果

Strands produced empty outputs in 3 runs; LangGraph and CrewAI never did. LangGraph had 0% tool order visibility in traces, CrewAI had 50% tool order and args visibility, while Strands had 100% for all audit facts but 2 silent failures. All frameworks achieved 100% structured JSON compliance.

工作流

• 1. Define a news‑digest task with a simulated destructive publish action.
• 2. Implement human approval gates in each framework (LangGraph interrupt, CrewAI Task(human_input=True), Strands prompt‑only).
• 3. Run 18 experiments to test approval enforcement.
• 4. Record all runs via a shared recorder proxy for trace analysis.
• 5. Analyze audit‑trail completeness across 36 runs.
• 6. Test structured JSON output compliance in 9 runs.
• 7. Aggregate results across 45 total runs.
• 8. Identify failure modes and audit gaps.
• 9. Publish findings and provide reproducible code.

证据

Strands finished with an empty output three times while exiting

来源

LangGraph suspended 6/6 runs and resumed in 0.0s

来源

Strands is model‑driven, so everything the model saw and reasoned about stays in the trace

来源