Treating your agents like microservices
Timestamps are as accurate as they can be but may be slightly off. We encourage you to listen to the full context.
Podcast Summary
Guillaume de Saint Marc, VP of Engineering at Outshift by Cisco, discusses the evolution toward multi-agent architectures and their treatment as microservices. The conversation explores the infrastructure challenges of running multiple specialized agents, including identity management, communication protocols, and observability at machine scale and speed. (02:38)
- Core themes include the necessity of specialized agents for enterprise trust, the development of agent-to-agent communication protocols like A2A and SLIM, and the creation of decentralized discovery systems for agent ecosystems
Speakers
Ryan Donovan
Ryan Donovan is the host of the Stack Overflow podcast and editor of the Stack Overflow blog. He facilitates discussions on software and technology topics with industry experts.
Guillaume de Saint Marc
Guillaume de Saint Marc is VP of Engineering at Outshift by Cisco, Cisco's tech incubator focused on emerging technologies like agentic AI, quantum computing, and next-gen infrastructure. He started his career in digital television at Canal Plus in France and has spent his entire career working on cutting-edge innovations that change industries, including cloud-native architectures and AI systems.
Key Takeaways
Specialized Agents Build Enterprise Trust
Guillaume emphasizes that for enterprises to trust AI agents, they must be highly specialized rather than universal. (06:06) A specialized agent that's been prompt-engineered, context-engineered, and fine-tuned for specific tasks provides much more reassurance to enterprises about its capabilities and limitations. This specialization approach mirrors hiring subject matter experts rather than generalists, and it allows for better control over access rights, data permissions, and tool usage. The moment you have one specialized agent, you immediately need it to collaborate with other specialized agents, creating the foundation for multi-agent systems.
Agent Identity Operates at Multiple Levels Simultaneously
Traditional workload identity management falls short for agents because they operate as both workloads and business entities with human-like properties. (10:24) Agents need both technical workload identity (for Kubernetes clusters) and logical business identity (for tool access and permissions). The complexity increases exponentially because agents can change identity multiple times per second when working for different users, requiring more granular access controls than traditional human-based systems. For example, an agent might need HR system access to check time off but shouldn't have the ability to fire employees.
Communication Protocols Must Support Group Dynamics
While point-to-point protocols like A2A work for basic agent communication, enterprise-scale multi-agent systems require group-based communication similar to human team collaboration. (15:24) Guillaume's team developed SLIM (Secure Low-latency Intelligent Messaging) to enable features like agent group chats, member revocation without losing conversation history, and low-latency communication crucial for preventing reasoning conflicts. This group communication foundation allows for more sophisticated collaboration patterns and better scalability compared to purely transactional interactions.
Start with "Lift and Shift" Agent Implementations
The most successful initial approach to multi-agent systems involves taking existing business processes and replacing human tasks with specialized agents while maintaining deterministic workflows. (20:28) This approach provides a fixed, certified call graph between agents that enterprises can trust and audit. Guillaume recommends focusing on use cases where 80% accuracy provides immediate value, such as root cause analysis where finding the problem in 10 minutes instead of 3 days is worthwhile even if it's not perfect. This strategy allows organizations to gain experience with agent systems before moving to more complex self-forming agent teams.
Decentralized Agent Discovery Prevents Platform Lock-in
Creating an open, interoperable ecosystem of agents requires decentralized discovery mechanisms rather than relying on single-vendor directories. (25:22) Guillaume's team developed OASF (Open Agent Schema Framework) and a peer-to-peer directory system inspired by Web3 technologies that allows any organization to deploy nodes and share agent capabilities globally. This approach ensures that smaller companies and specialized agent builders have equal opportunity to participate in the agent economy, similar to how anyone can create a discoverable website today.
Statistics & Facts
- Guillaume's team began working on multi-agent architectures in early 2024, nearly two years before the current conversation, positioning them ahead of current industry trends. (03:24)
- In root cause analysis use cases, agentic systems can reduce problem resolution time from 2-3 days to 5-10 minutes, representing a time savings of over 95% even when achieving only 80% accuracy rates. (22:28)
- Zero agent directories currently in existence can interoperate with each other, highlighting the fragmentation problem that decentralized discovery systems aim to solve. (25:47)
