Architecting Agentic Systems at Navan

The Breakthrough

Navan manages agentic complexity by treating skills as the primary unit of context, allowing a single master agent to dynamically load and test pluggable capabilities rather than relying on complex multi-agent orchestration.

What Actually Worked

• Hook-based Observability: Instead of relying on standard logs, the team uses hooks to intercept pre-tool and post-tool calls. These hooks emit structured traces containing the current goal, reasoning, belief status, and tool-call confidence scores.
• Trajectory-based Testing: Because agents are nondeterministic, the team evaluates performance by scoring the trajectory of an agent from start to goal rather than asserting specific output values.
• Dynamic Context Loading: The system uses progressive disclosure to manage context windows, starting with a limited scope and expanding metadata only as required by the specific skill being executed.
• Fine-grained Guardrails: Authorization is enforced at the tool-call level rather than the edge, allowing the system to distinguish between an agent acting on behalf of a user versus using a service account.

Context

Navan, a travel and expense management platform, faces the challenge of building reliable agentic workflows in production. The speakers argue that the industry is currently in a phase similar to the 2015 microservices boom, where the temptation to over-engineer complex systems often outweighs the benefits. They emphasize that operational maturity in agentic systems requires moving away from traditional stateless API patterns toward stateful, persistent sessions that can handle complex, multi-step reasoning chains.

Content References

• tool: BrainTrust, context: mentioned
• tool: Model Context Protocol (MCP), context: mentioned
• tool: OpenTelemetry (OTel), context: mentioned