The Agentic Architecture: Skills over Complexity
Drawing parallels to the 2015 microservices boom, the speakers argue against premature complexity. Just as a well-structured monolith is often superior to a distributed system, a single, well-defined agentic loop is more reliable than a complex multi-agent orchestration. Navan utilizes a "master agent" pattern that dynamically loads "skills." These skills serve as the unit of context—pluggable, testable, and reusable components that encapsulate both instructions and tool execution. This approach allows for progressive disclosure of context, keeping the agent focused by only loading relevant metadata as needed.
Operationalizing Non-Deterministic Systems
Traditional logging fails when agents generate massive amounts of reasoning. Instead, Navan implements hooks that intercept every tool call to emit structured traces containing the goal, reasoning, and a confidence score. This allows for "human-in-the-loop" intervention when the agent's confidence is low or when an answer is inferred rather than retrieved.
Testing nondeterministic systems requires a shift from asserting static outputs to scoring trajectories. By evaluating the path an agent takes from start to goal, teams can measure efficiency and completeness. This trajectory-based evaluation helps identify regressions in agent behavior that standard unit tests would miss.
Governance and Authorization
In an enterprise context, the line between user action and agent action is blurred. When an agent acts on behalf of a user (e.g., booking a flight based on a price trigger), traditional identity models break down. Navan enforces governance through guardrails that run before and after every tool call. This ensures that authorization decisions are made at the point of execution rather than at the network edge, providing fine-grained control over what an agent can do with a user's service account or permissions.
The State of the Industry
- Solved: Runtime environments and basic tool-calling protocols (like MCP) are maturing, with cloud providers offering standardized support.
- Emerging: Observability (adapting OpenTelemetry for agentic flows) and agent-to-agent communication protocols are still in their infancy.
- Unsolved: Cost management remains a significant challenge, as current vendor incentives favor high token consumption. Debugging and replaying agentic failures also remain difficult, though the speakers suggest that using AI to analyze agent traces may eventually mitigate this cognitive overload.