Agents Are the New Microservices: Lessons from Navan's AI…

Explore how AI agents mirror microservices in 2015, Navan's master-agent model, and the challenges of testing, logging, and cost.

Ever wonder why AI agents feel like the next big shift, just like microservices did a decade ago? In this clip, Navan’s engineers explain the parallel.

⚡ TL;DR: Agents are the new microservices, Navan uses a master agent that loads pluggable skills, and testing them requires scoring whole reasoning paths rather than simple output checks.

The Simple Version

Think of an AI agent as a smart travel assistant that can decide when to book a flight on its own. It blends decision‑making with tool use, much like a modern app blends services.

How It Actually Works

Agent Architecture

Navan runs a single master agent that gradually loads skills. Each skill is a reusable piece of logic that carries its own context.

🎯 Think of it this way: A skill is like a Lego block that you can snap into different models without rebuilding the whole structure.

Guards and Hooks

Guardrails sit before and after every tool call. Hooks capture the goal, reasoning, and a confidence score for each step.

🎯 Think of it this way: It's like a security camera that watches every door you walk through, logging who entered and why.

Testing Nondeterminism

Because agents produce varied outputs, you score entire reasoning pathways instead of asserting a single answer.

🎯 Think of it this way: Rather than checking a single roll of a die, you rate the whole sequence of rolls for fairness.

Technical Deep Dive

Memory: Agents move from simple RAG retrieval to episodic memory, letting them recall past interactions as events.

Cost: Running many tool calls and evaluating trajectories still lacks a reliable cost model, making budgeting hard.

Replay & Standards: Replaying agent behavior for debugging is emerging, and early standards are shaping how logs should be structured.

Why This Matters

Agents change how we think about authorization and ownership in code. When an AI can act on its own, old notions of a user or service principal break down.

Understanding the master‑agent pattern helps teams design more modular, testable AI systems, but it also highlights open problems around cost and reliability.

Key Takeaways

Agents mirror the microservices shift of 2015 — modular, reusable, and prone to over‑engineering.
Navan’s master agent loads skills that act as context‑rich building blocks.
Guardrails and hooks provide visibility into each reasoning step.
Testing agents means scoring whole trajectories, not single outputs.
Memory evolved from simple retrieval to episodic recall.
Cost and reliable replay mechanisms remain unsolved challenges.
Future standards will likely dictate how agent logs are structured and shared.

Agents Are the New Microservices: Lessons from Navan's AI Architecture