新增第二個 Agent 前應先確定哪些事項

A second agent brings failure modes a single agent never had. The Chain of Thought collection on this theme, many agents in one system, is four conversations. I would take one decision from each, in this order.

Which orchestration shape can your data support?

Siva Surendira, CEO of Lyzr, walks through the shapes in Why Enterprises Need a Different Approach to AI Agents (30 Apr 2025). He names managerial flows, DAG flows, and a hybrid that combines a manager with a DAG.

Each shape runs into the enterprise's integration limits and its data readiness. Siva says 95% of all AI work today done by enterprises remains a proof of concept and never reaches production. The episode summary describes agentic AI as having exploded in 2025, and it frames the open problem as moving past prototypes to agents that stay reliable and valuable at scale. He frames Lyzr around agent frameworks enterprise developers can use and trust.

I would list every system the agents must call and every dataset they must read before I name a shape. An unfinished list stalls a manager, a DAG, and a hybrid in the same place.

What should you measure once a second agent exists?

João Moura, founder and CEO of CrewAI, takes up that risk in The Emerging AI Agent Stack (25 Jun 2025). João maps the orchestration layer everyone is racing to standardize. He is blunt about how many more failure modes you inherit the moment there is more than one agent.

The episode summary says scaling multi-agent systems for knowledge work requires measurability, reliability, and trust. Once enterprises move past initial curiosity, the questions are provisioning, authentication, and measurement for hundreds or thousands of agents in production. I would write those down while the agent count is still small: who may provision an agent, how that agent authenticates, and what you measure when one agent hands work to another. Standardizing orchestration does not settle those operating questions.

How do you apply engineering discipline when nothing is deterministic?

Mikiko Chandrasekhar, then a staff developer advocate at MongoDB, makes this case in Mastering Multi-Agent Systems (23 Jul 2025). Treat agents as software products and apply engineering discipline. That is what keeps a multi-agent system reliable when nothing inside it is deterministic. The summary calls mastering that reliability the ultimate challenge if agentic systems are going to work in production.

Her background spans the data-to-AI pipeline. She describes MongoDB's vision as the memory store for agents, from data storage and vector search through debugging chat logs. I would spec that store before agents share state: a place for the data, a way to search it, and logs you can open when a run goes wrong. A later run may differ from this one, so those logs are how you debug a system that is not deterministic.

What has to exist before one agent calls another?

Giovanna Carofiglio, Distinguished Engineer and Senior Director at Outshift by Cisco, covers the Internet of Agents in Architecting Reliable Agentic AI (9 Jul 2025). The episode was recorded at Cisco's office in San Jose. She is a leader of the AGNTCY Collective, an open-source initiative that Cisco, LangChain and other founding members launched in March. She outlines agents that collaborate across the enterprise and the internet. The summary says that outcome needs standards, protocols, and evaluation methods.

The collection page gives the builder three jobs: agent discovery, the transport layer under agent-to-agent traffic, and a clear account of why MCP and A2A solve different problems. I would decide whether this connection is discovery, transport, or a protocol choice, then use her account of how MCP and A2A differ before binding the system to either one.

Checklist

• Write the integration points and the data each step needs, then choose managerial, DAG, or hybrid.
• Assume more failure modes as soon as a second agent exists. Define provisioning, authentication, and measurement early.
• Treat the agents as a software product, with engineering discipline and a memory store you can search and debug.
• Plan discovery and the transport under agent-to-agent traffic as their own work, and treat MCP and A2A as different problems.
• Judge the design by whether it can leave proof of concept status. Siva puts that miss rate at 95% of all AI work done by enterprises today.

More on this topic, with the related episodes, is on Chain of Thought. It draws on this episode.

Subscribe to the Chain of Thought newsletter for new episodes and write-ups like this one.

Drafted with AI assistance from the episode transcripts.