Deciding what not to expose

The idea in brief

Much cloud infrastructure assumes a person who can read documentation, write configuration and debug when something breaks. A Latent Space episode on agent experience argues that agents don't have that context. The same question applies to people who use a platform without being infrastructure specialists.

Points from the Latent Space episode

In "Why AI Infrastructure must evolve for Agent Experience" (July 2026), Akshat Bubna, CTO of Modal, talks with swyx and Vibhu. Points from the episode:

• Traditional cloud tooling assumed a human operator who carries implicit context about how it works.
• Modal shifted part of its SDK work from developer experience to agent experience, and argues the same improvements help both.
• Defining infrastructure in code rather than configuration files makes it easier for agents to change and observe their environments.
• Guardrails for agent sandboxes should rely on hard boundaries, not only permissions mediated by a model.

A thin layer over what exists

The CNCF Platforms White Paper recommends building the thinnest reasonable layer over existing managed services, and offering capabilities through self-service APIs, CLIs and templates. That is a way of exposing a smaller surface without rebuilding what is underneath.

The balance

Hiding detail and hiding evidence are different things. A user may not need to know about clusters, identity plumbing or gateway policies, but still needs enough signal to tell where a problem is. The post on abstractions and evidence looks at that side.

@article{adam2026whatnottoexpose,
  title   = {Deciding what not to expose},
  author  = {Adam, Ammar},
  journal = {www.ammaradam.dev},
  year    = {2026},
  month   = {Sep},
  url     = {https://www.ammaradam.dev/posts/what-not-to-expose/}
}