Pulumi shipped a Context API that turns infrastructure state into a queryable graph, and the design premise is that the useful questions are all about relationships. Not what resources exist but which of them a given database is reachable from, what a change would touch downstream, which stacks have no policy coverage. “The Context API connects this data so you can ask questions that depend on the relationships.” The query surface is deliberately small: “a query is a JSON document with a handful of clauses” — anchor to pick a starting node, traverse to walk edges, return to say what comes back — with the vocabulary itself discoverable through pulumi api GetGraphQuerySchema. It is in public preview for Enterprise and Business Critical organizations, needs CLI v3.243.0 or later and a resources:search permission.
What makes it notable is that it is agent-first without being an agent product. The intended callers are Pulumi Neo and third-party coding agents — Claude Code, Cursor, Codex — running authenticated CLI commands, which means the credential model, the permission scope and the audit trail are the ones the organisation already has. That is a more conservative architecture than exposing a new agent endpoint, and it is why a JSON query language beats natural language here: a query an agent composes is one a human can read back before it runs. The honest caveat is stated rather than buried: “during public preview, the graph vocabulary and limits may change,” which is the correct warning for anything an agent might hard-code against.
The catalog carries 35 API pages for Pulumi, and the Context API arrives into a surface already shaped for this. The Pulumi Insights API is the resource-search ancestor of a graph query, the Neo API and AI Agents API are the first-party agent surfaces the graph feeds, the Environments API holds the configuration and secrets those agents resolve against, and the AuditLogs API is what makes an agent-issued query reconstructable afterwards. The Automation API remains the write path — the graph is for asking, not changing.
Pulumi scores 42.8, developing on the Kin Score, with access and commercial clarity at 60.5 against contract governance at 0.0. Agent Readiness is 27.6, agent-aware, with agent_skills and reversibility_documented lit — the second of which matters more than it sounds for an infrastructure provider, since the whole value proposition is that a change can be undone. mcp_server is unlit, and that is not an oversight so much as a road not taken: Pulumi’s agent door is the authenticated CLI, the same one a human uses, rather than a separate protocol endpoint. It is a defensible choice and it costs them a dimension. What it buys is that nothing about an agent’s access is novel — the permission it needs, resources:search, is a permission an administrator already knows how to reason about.