Anthropic just cut off third-party agent tools, OpenClaw and others, from running persistent agents on Claude subscriptions. They called it capacity management. Builders called it something else. The message is clear: agentic use is a separate tier now, and the labs intend to own the runtime, not just the inference.

The three biggest labs, Anthropic, OpenAI, and Google, Are now formally sharing intelligence to detect Chinese competitors systematically querying their models to train knockoffs. First coordinated IP-defense operation across competing rivals.

Meanwhile, Google dropped Gemma 4 under Apache 2.0, 31B parameters, multimodal, 256K context, commercial use unrestricted. For anyone running local inference on capable hardware, this is the first serious frontier-grade option with no strings attached.

AGENDA

10:00 – 11:00 Agent access tiers: who controls the runtime
11:00 – 11:10 Community Announcements
11:10 – 12:00 Geopolitics and the API layer
12:00 – 1:00 Open models vs. hosted APIs: does Gemma 4 change the math?

WHAT WE'LL DISCUSS

Anthropic says agent lockdowns are capacity management Do you believe them? and what does it mean if they're building a walled garden around agentic use?
The labs are now coordinating like a cartel to protect model IP from China, is that good security or the beginning of something else?
Gemma 4 is Apache 2.0 and frontier-class: Does this finally make self-hosting the right call for serious builders?

HOW TO FIND US

Armistice Coffee, Roosevelt neighborhood. Breezeway out back, just look for the AI Northwest banners.

WHO SHOULD COME

Builders running agent workflows, founders with federal or enterprise clients, and anyone whose stack touches an API that got more complicated this week. If you've wondered whether to own your models or rent them, come with opinions.

Free — buy your own coffee.

STAY IN TOUCH

ainorthwest.org

community.ainorthwest.org

linkedin.com/company/ai-northwest

AI Northwest Coffee Social - Agent Lockout, Distilling Cartels, and Gemma 4 Hits the Wires