The Brief · Issue 4

Week of 20 July 2026 — Snowflake goes agentic, DataRobot bets on sovereignty

Snowflake ships an agentic control plane and CoCo GA, DataRobot leans into sovereign AI, and the EU's AI Act amendments hit the Official Journal.

September 22, 2026

Compiled on 22 September 2026 for The Brief's launch archive, from sources published between 20 and 26 July 2026.

Four stories that mattered

Snowflake makes CoCo general and adds per-user AI cost caps

At its July AI Pulse event on 21 July 2026, Snowflake announced that CoCo — its AI-agent workspace, spanning both a desktop app and cloud-hosted agents — reached general availability, alongside a self-governing agentic control plane across the platform. Companion releases included Cortex Agent Sharing, which lets teams share configured agents the way they already share data; per-user AI cost quotas that let admins cap an individual's spend rather than only a workspace total; a direct integration linking Snowflake's customer context to Meta ad campaigns; and an online machine-learning feature-serving layer for real-time scoring.

The per-user cost quotas are the detail worth testing first: they're a direct answer to the "who burned the AI budget" question that a shared workspace quota can't answer on its own.

Source: Snowflake, AI Pulse – July 2026 Product Announcements

DataRobot pitches "run anywhere" governance as sovereign AI

DataRobot said on 22 July 2026 that its Agent Workforce Platform now applies the same governance, monitoring and control whether deployed in the public cloud, on-premise, air-gapped, in a virtual private cloud, or split across multiple clouds — letting customers bring their own models, including open-source ones, and run them entirely within infrastructure they control. The company framed the move around what it called a shift of sovereign AI from a government concern to a boardroom one, citing an incident in which a government directive cut a widely used AI model's access overnight with no transition period.

For regulated-industry analytics teams — financial services, defence, healthcare — the pitch is really about optionality: the ability to move a model off a public cloud without rebuilding the governance layer underneath it.

Source: DataRobot, DataRobot Gives Enterprises Full Control Over Where and How Their AI Runs

The EU's AI Act amendments become law

The EU's Digital Omnibus on AI — the package the Council approved on 29 June (see our 6 July issue) — was published in the Official Journal of the European Union as Regulation (EU) 2026/1744 on 24 July 2026, and entered into force three days later. Publication starts the clock on the amended deadlines: 2 December 2027 for high-risk standalone AI systems, 2 August 2028 for high-risk AI embedded in regulated products, and 2 December 2026 for the Article 50 watermarking grace period.

What was a "should we start preparing" question a month ago is now a dated legal obligation; teams tracking AI Act exposure should treat 24 July as the reference date for their compliance calendar, not the original June adoption.

Source: NicFab, Digital Omnibus on AI: Regulation (EU) 2026/1744 Is Published in the Official Journal

Tool moves

One how-to

Set a per-user AI spend guardrail, not just a team one. Snowflake's new per-user cost quotas are a reminder that most AI budget overruns aren't a team problem until they are — they start with one person's workflow.

  • Find out whether your data platform and any LLM gateway you use support quotas at the individual level, not only at the workspace or project level.
  • Set a default quota low enough to catch a runaway agent loop, and a documented, fast path to raise it for people who legitimately need more.
  • Alert on trend, not just threshold: a user who doubles their weekly usage two weeks running is worth a conversation before they hit a hard cap.
  • Treat the quota as a guardrail metric, not a target — the goal is catching runaway cost early, not making everyone use exactly their allotment.
  • Review quotas quarterly against actual usage; a limit set once at rollout is usually wrong within a few months in one direction or the other.

This is a small piece of a larger discipline — guardrails around LLM usage matter as much for cost as they do for output quality, and both fall under the same data governance umbrella.

One number

628,000 — the seasonally adjusted annual rate of new single-family home sales in the US for June 2026, released jointly by the Census Bureau and the Department of Housing and Urban Development on 24 July 2026, up 1.6% from May's revised 618,000 but 5.6% below June 2025.

What it doesn't tell you: new-home sales are a small, volatile slice of the housing market — existing-home sales account for the large majority of transactions and aren't included here — and the Census Bureau's own estimate carries a wide margin of error, so a single month's rise can reverse in the next revision.

Source: U.S. Census Bureau and Department of Housing and Urban Development, Monthly New Residential Sales, June 2026

Get the next one