Sign InOpen Brain
AI EngineerVideoSource Linked

Agents Are Where Microservices Were in 2015 — Roberto Milev & Uday Kanagala, Navan

Navan treats production agents as stateful services needing persistence, composable skills, traces, authorization gates, and trajectory-based tests before multi-agent complexity.

AI Engineer · Aug 29, 2026
Open Source Open MarkdownOpen JSON
Source Summary

Navan’s emerging stack covers runtime, memory, context, operations, and orchestration. It persists and rehydrates sessions, composes context from independently testable **skills**, and places guardrails **before and after every tool call**.

Practical Implication

Start with one reliable agent and progressively load domain skills instead of defaulting to a multi-agent graph. Capture goals, tool calls, inferred answers, and execution trajectories so failures can become test cases and policy changes.

Agent-Ready Context
Navan’s emerging stack covers runtime, memory, context, operations, and orchestration. It persists and rehydrates sessions, composes context from independently testable **skills**, and places guardrails **before and after every tool call**.

Start with one reliable agent and progressively load domain skills instead of defaulting to a multi-agent graph. Capture goals, tool calls, inferred answers, and execution trajectories so failures can become test cases and policy changes.

Cloud runtimes and tool protocols are converging, but cost prediction, replay, agent observability, and orchestration remain unsettled. Navan’s single-master pattern reflects its use cases rather than a universal architecture.
Connected Context · Feed7 Judgment

This reinforces a single-agent-first architecture in which independently testable skills, persistent sessions, tool-boundary guardrails, and captured trajectories form the operating stack. It makes multi-agent orchestration an earned escalation rather than a default, while leaving replay, observability, and cost prediction open. The pattern is credible as Navan’s implementation choice but not established as universal.

SafeEvolve: Harness-Policy Co-Evolution from Agent Experience for Safety AlignmentNavan proposes turning captured failures into tests and policy changes; SafeEvolve supplies a more explicit mechanism for deriving reversible harness updates and policy training from those trajectories.Tribal Dungeons of Global Shipping: AI Agents at Global Scale — Dmitry Buykin, MaerskMaersk provides large-scale operating evidence for Navan’s emphasis on bounded tools, replayable trajectories, and continuously maintained procedures, while exposing the expert labor such a stack can still require.The Era of Compound Engineering — Kieran Klaassen, Every/CoraBoth treat reviewed experience as reusable infrastructure: Navan captures trajectories and composes testable skills, while compound engineering stresses that corrections and judgments must be curated into context for later runs.Don't Build Agents You Can't Answer For — Addy OsmaniNavan’s capture of goals, calls, inferred answers, and trajectories supplies the evidence base needed for the candidate’s requirement that agent work remain explainable, testable, and owned before shipping.
Context Map
agentcoding#harness-engineering#skills#agent-reliability
Uncertainty
Cloud runtimes and tool protocols are converging, but cost prediction, replay, agent observability, and orchestration remain unsettled. Navan’s single-master pattern reflects its use cases rather than a universal architecture.