← back to the edition
AI ENGINEERING 3 min read

"Agent Clouds": the next infra fight is where your agents run

Once the models commoditize, the moat moves to who holds your agents’ context, state, and spend.

primary source
What happened

On Latent Space, Databricks co-founders Matei Zaharia and Reynold Xin pitched the company as "the operating system for enterprise agents" via Omnigent, an open-source meta-harness that runs agents across Claude Code, Codex, and Cursor. Their five named problems: portability across harnesses, collaboration/session history, security via stateful policies, spend controls, and common APIs — with governed access to live business data, not just telemetry.

Why it matters

This is the opening shot of a platform war over where agents run and what they’re allowed to touch. If it works, you stop hand-rolling permissions, memory, and cost caps and rent them as a layer. If it’s overreach, you’ve handed your operational context — the actual moat — to a vendor.

Eduardo's take

Honest reaction: is Agent Cloud real, or PaaS with new branding? I operate client agents on VPS and EC2 by hand — SSH, systemd, env files, a Redis for state. Databricks’ five problems are real; I hit all of them. But "the OS for enterprise agents" is the Heroku promise again. The genuinely new part worth paying for is governed access to live operational data — which is exactly what my MCP servers already do. So I read this less as "a new cloud" and more as "MCP-shaped data governance, productized."

Source: Latent Space — "The Agent Cloud" w/ Matei Zaharia & Reynold Xin (Jun 24, 2026)

EC TV is written by Eduardo Cruz — a senior Laravel engineer who ships production AI agents and MCP servers.

Work with me Read the deep-dives
← back to the edition