PECS: 基于 LangGraph 的四角色多智能体任务求解框架(Planner/Executor/Critic/Synthesizer)。WebShop 真实环境 +25pp (25% vs 0%);GAIA 官方 53 题 26.4% vs ReAct 24.5%(McNemar 不显著)。含 AST 沙箱、50000 token 硬预算、FastAPI 限流/混沌/CI/Prometheus。
python flask multi-agent ai-agents cost-optimization ablation-study ai-agent llm-agent langgraph deepseek-api token-optimization gaia-benchmark plan-execute-reflect agentbench ast-sandbox
-
Updated
Sep 18, 2026 - Python