Architects Debate Separating Governance from Agent Logic

- Open-source frameworks like Microsoft's AutoGen and CrewAI are explicitly designed to support multi-agent collaboration, providing structures for defining agent roles and managing their interactions. AutoGen, in particular, emphasizes a multi-agent conversation framework, while CrewAI focuses on orchestrating role-playing agents for collaborative tasks. - Architectural patterns for multi-agent systems are becoming more formalized, with Google outlining eight key patterns including sequential pipelines, parallel fan-out/gather, and human-in-the-loop architectures. These patterns serve as building blocks, allowing developers to create more reliable and scalable systems by assigning specialized roles to individual agents. - The separation of concerns is a driving principle, where decoupling the agent's workflow logic from its inference or search strategy can reduce technical debt and improve governance. This allows for independent optimization and makes the system easier to audit and test. - In China, major tech companies like Tencent and Alibaba are deploying multi-agent systems at a massive scale within their ecosystems. Tencent's Agent Runtime, for instance, handles billions of tool calls daily within WeChat, showcasing real-world application of agentic orchestration. - From a security and governance perspective, the shift to multi-agent systems introduces new risks, as failures can emerge from agent interactions rather than just individual model errors. This has led to research into concepts like a "Simulation Governance Kernel" to ensure deterministic execution and auditable state changes, as well as Policy-as-Code (PaC) architectures to enforce runtime rules. - China's regulatory landscape for AI is moving towards a "Granular and Vertical" strategy, with specific rules for different AI applications, contrasting with the EU's broader risk-based approach. Recent draft rules from the Cyberspace Administration of China (CAC) specifically target "anthropomorphic interactive services," indicating a focus on the user-facing aspects of agentic AI. - When comparing orchestration frameworks, LangGraph is noted for its ability to build controllable, stateful agents, making it suitable for complex, deterministic workflows, while AutoGen is preferred for more flexible, conversational collaborations. LangChain provides modular pipelines and strong integrations for Retrieval-Augmented Generation (RAG). - A key challenge in multi-agent systems is managing state and preventing independent state mutations that break determinism and make operations impossible to replay for debugging. This has prompted architectural proposals for a central governance layer to arbitrate, order, and schedule actions before committing state changes.