Milev compares today's agent enthusiasm with the early microservices wave. Useful infrastructure eventually emerged from microservices, but teams also learned that distributing a poorly structured system only multiplies its problems. The same lesson applies to agents: if a team cannot make one agentic loop reliable, multi-agent orchestration adds boundaries and failure modes before the foundation works.
Production agents differ from conventional stateless services because they need persistent sessions, isolation and a longer lifecycle. Navan combines a managed runtime with its own session persistence, then builds memory from short conversations into longer-term and episodic records. Context is assembled progressively from skills that package instructions and tool execution into reusable, independently testable units.
Traditional logs are not enough for nondeterministic flows. Hooks around tool calls capture goals, reasoning signals and confidence, while evaluation scores trajectories rather than requiring one exact output. Security is equally contextual: an agent may act for a user or through a service account, so fine-grained policy checks and guardrails run before and after tool calls to decide whether an action is allowed.
The speakers see runtime, memory and tool protocols becoming more standardized, while observability, testing, cost control and replay remain less mature. Their practical architecture favors one master agent that progressively loads skills and sub-agents only when necessary. Team boundaries can later use agent-to-agent contracts, but complexity should arrive after the single-agent path is understood. Closing thanks and applause are omitted.
Watch on YouTube
