Summary
Roberto Milev, Chief Architect at travel and expense management company Navan, and Uday Kanagala from the same architecture team draw a compelling historical parallel: AI agents today are roughly where microservices were in 2015 — full of promise, but lacking the accumulated tooling and patterns that make them reliable at scale. Their talk maps out the layered reference architecture that has crystallized from running many agents and millions of tokens per day in production.
The stack they describe has four principal layers. The runtime layer addresses the fundamental statefulness of agents, contrasting with the stateless services engineers are used to building — Navan runs on AWS Agent Core but has built custom session persistence and rehydration on top. The memory layer evolves from basic RAG through short-term conversational memory to long-term and episodic memory. Context management is handled by treating ‘skills’ as composable units — each skill bundles both domain instructions and tool-execution logic, enabling dynamic context assembly with progressive disclosure. The observability layer emits structured traces capturing an agent’s current goal, reasoning, belief state, tool calls, and a confidence score, allowing engineers and agents to jointly triage failures.
A significant portion of the talk addresses testing non-deterministic systems, where the speakers candidly admit they are still working out answers. The microservices analogy proves useful here too: just as ‘if you can’t build a monolith, don’t build microservices,’ the equivalent lesson for agents is to master a single agentic loop before attempting multi-agent orchestration.
📺 Source: AI Engineer · Published August 29, 2026
🏷️ Format: Deep Dive
