When Washington can switch off your model: building for the AGI-governance era
The US executive branch already forced Anthropic to cut access to a Claude model, and is now gating GPT-5.6 customer-by-customer. If your product rides one closed model, that’s a single point of failure with a political trigger.
Nathan Lambert argues we’ve crossed into "the AGI era of AI governance," dated to the executive branch forcing Anthropic to turn off access to a Claude 5 model after Amazon flagged cybersecurity risk. Days later, The Rundown reported the administration asked OpenAI to stage GPT-5.6’s release to government-approved partners only — access granted "customer by customer." Lambert’s warning: releases are now judged on vibes by an executive branch with minimal technical talent, and a similar move against an open model is plausible within months.
This reframes provider risk. We design for a model being deprecated, rate-limited, or price-hiked — not for it being switched off by a government between sprints, or for our account never getting approved. "Customer by customer" access means your competitor ships on the new model while you wait. The hedge is no longer optional: a provider abstraction layer, evals that run against more than one backend, and a tested fallback to an open weight you can self-host.
I sell consulting and Throughline on top of closed frontier models — so this is my exposure, not a thought experiment. Never let a client’s core flow depend on a model id I can’t swap: a provider interface, prompts and evals not tuned to one vendor’s quirks, and a documented degraded-mode on an open weight. Continuity isn’t a feature you bolt on after Washington makes the call; it’s an architecture decision you make now, while everything still works.
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