Technological Evolution of AI Agents Likened to Microservices in 2015
Navan architects Roberto Milev and Uday Kanagala state that the operational infrastructure of AI agents is undergoing a maturation process similar to microservice architectures in 2015.
Navan Chief Architect Roberto Milev and Architect Uday Kanagala argue that the technology for running AI agents is currently settling in, much like microservices did in 2015.
Current State of Agent Infrastructure
AI agents stand out as stateful, difficult to observe, and costly systems.
It is emphasized that those who cannot properly establish a single agent loop should not build multi-agent systems.
Memory Management and New Approaches
Memory management is now evolving beyond the limits of RAG into a structure where context is treated as skills.
Providers integrate this process with semantic search, short-term chat, and episodic memory with long-term memory.
Observability and Debugging
It is stated that traditional logs fall short for debugging in long-running agent processes.
Intersections at decision points and tool call traces are effectively used to manage human intervention.
Enterprise Governance and Cost Challenges
For enterprise-grade regulation and orchestration, single-agent discipline must be established first.
Making systems debuggable and billable after a rapid prototype creates a significant cost challenge.