Cloud Agent Primitives

zach-lloyd's inventory of what you actually need to run an agent fleet off-laptop:

1. Environments — defined tool-chain + code access per agent.
2. Hosting — Daytona, E2B, Docker sandboxes, namespace (or your own compute).
3. Visibility — watch what the agents are doing while they run.
4. Human-in-the-loop handoff — "if an agent is working in a cloud system and it can get only 80% of the way there, and you want an engineer to pick up what it did, that should be easy."
5. Programmability — API/SDK/CLI launch; configurable environments; pull artifacts and conversations back.

Why this matters

This is a crisp procurement checklist for 2026 agent platforms. Notice the overlap with Niels Bantilan's durability layer (durable-observable-debuggable-agents) — Lloyd is basically saying "everything you need in the user interface" while Bantilan says "everything you need in the runtime." Operators building in-house should match these lists against their chosen platform.