Agents Are Where Microservices Were in 2015 — Roberto Milev & Uday Kanagala, Navan
AI Engineer World's Fair 2026 · 19:27
Business travel, payments, and expense management
Navan combines business travel booking, payments, and expense management for employees, travel managers, and finance teams. Travelers can find flights and hotels with 24/7 support, while Navan Expense categorizes transactions, flags spending outside company policies, and syncs expense data with accounting and ERP systems. Companies can issue Navan corporate cards or use Navan Connect to link eligible existing cards to the expense workflow.
Ariel Cohen and Ilan Twig founded the company as TripActions in 2015; Cohen remains CEO and Twig CTO. It adopted the Navan name in 2023 as it consolidated travel, corporate card, and expense offerings. Its Ava virtual travel agent uses Navan Cognition, a modular framework coordinating specialized agents for retrieval, transactions, and policy checks. Supervisory flows validate outputs and API calls; for high-stakes actions, language models extract parameters while separate fulfillment checks govern execution.
Navan acquired Reed & MacKay in 2021, adding premium travel management and expanding into Europe, and began trading on Nasdaq as NAVN in 2025. Its Navan Edge assistant brings personalized planning for flights, hotels, and loyalty benefits into a conversational interface. An OpenTable integration announced in 2026 adds restaurant discovery and reservation booking, using travelers’ location, itinerary, and preferences to connect dining plans with their trips.
Navan’s supplied recording archive contains one talk by Roberto Milev and Uday Kanagala. Use the paths below to explore its architecture, operational practices, and authorization questions. These descriptions reflect what the speakers reported in the recording; they do not establish Navan’s current infrastructure or product capabilities.
For the overall design, watch Agents Are Where Microservices Were in 2015. The speakers recommend a well-structured single agentic loop before introducing multi-agent complexity. Their discussion connects stateful runtimes, session persistence and rehydration, short-term conversation memory, long-term semantic memory, and episodic records. Follow the skill concept—domain instructions plus tool execution—as their reusable, independently testable unit for progressively loading capabilities.
Return to the same recording for the speakers’ account of operating nondeterministic systems. They describe hooks around sessions and tool calls that capture goals, reasons, confidence signals, and tool activity, and report using Braintrust for traces. This path is useful for exploring how inferred answers may prompt human guidance and how trajectory evals can assess progress toward a goal without demanding one fixed execution sequence.
Read the talk’s authorization discussion through its hypothetical delayed flight purchase: who bears responsibility when an agent acts later on a user’s behalf? The speakers describe fine-grained guardrails before and after tool calls and suggest A2A may help across organizational boundaries. Their assessment of observability, replay, debugging, standards, and predictable costs as developing areas is a recorded viewpoint, not a verified statement about today’s tools.
AI Engineer World's Fair 2026 · 19:27
Affiliations reflect their AIE appearances, not necessarily current employment.