Kamalakannan Nandagopal describes the limits of coding agentsAn AI coding agent is a tool-using AI system that can inspect, modify, and validate software within a repository. working inside one repository when real applications span many servicesA distributed system is a group of networked components that coordinate to provide one service despite separate state and partial failures. and environments. Postman's approach maps APIs, their implementation locations, consumers and downstream data flows into a shared context graphContext engineering designs the information, instructions, memory, and tool state an AI receives so it can perform a task reliably. grounded in code or production traces.
The team evaluates the graphAgent evaluation tests whether an AI agent completes tasks correctly, consistently, and within its required boundaries. using historical pull requests and design decisions, comparing its agent with generic code search and a graph-backed skill. Examples cover finding an appropriate API, understanding different consumer input shapes and assessing the impact of an API change. The reported outcomes are the team's findings, not independently verified benchmarks.
One evaluation fails because new consumers appeared after the graph was populated. Kamalakannan Nandagopal stresses that additions and removals must stay synchronized with implementation evidence or developers lose trust. He also describes architecture reviews that surface dependency cycles and telemetry gaps, and proposes adding business semantics and design rationale as further context.
Watch on YouTube
