After the Kill Switch: What the Trench Kept

Two weeks after Anthropic blacked out Fable 5 and Mythos 5, the interesting question was not whether the letter was fair. I already wrote that it was a sovereignty move, not a product incident. The interesting question was the one Slack kept asking in rooms nobody screenshots: what do we still have on disk?

That is the hangover. Executives processed the recall as a vendor event. Operators processed it as an inventory. If your "AI capability" lived in a SKU, you had an apology and a degraded tier. If it lived in evals, routes, weights, and folders the team owned, Monday still happened. This piece is that inventory, what the trench actually kept when the frontier went dark.

Two weeks of observed behaviour

I watched three patterns, over and over, in labs I trust and in holding-company teams that will never be named in a case study.

Procurement did not flinch. The invoice still said Anthropic, OpenAI, Google. Europe still wanted the stamp. Boards still wanted the rider. That is Market A. It did not learn. It waited for access to come back, slightly nerfed, and called the wait "responsible."

The trench dual-homed overnight. Qwen, Llama derivatives, DeepSeek forks, a Mac Studio that had been a joke in Q1, vLLM on a rented GPU, a router that used to be a convenience and was now the product. Not ideology. The demo was Tuesday.

Legal froze the people who had been shipping. The worst rooms were not the ones that switched weights. They were the ones that sent a memo: pause all generative tools until further notice. Shadow use did not pause. It moved onto personal accounts with worse logging. You taught the change-makers that the company is the incident.

What survived (write this down)

When the SKU died, four classes of artifact still ran. If you cannot point at these in your repo, you did not have an AI programme. You had a subscription.

• Task evals tied to real work. Ticket triage, brief extraction, JSON that must parse, a brand check that fails on a forbidden phrase. Small. Ugly. Owned. Leaderboard cosplay did not help anyone on Monday.
• A model router. Named routes, fallbacks, cost ceilings. When Fable disappeared, the teams with a router changed a string. The teams with "we use Claude" in a slide opened a war room.
• Weights you could still load. Local, VPC, a box in a closet. I am not romantic about the closet. I am describing discontinuity insurance, the same hedge as open source as a hedge.
• Folders the agent already knew how to open. Skills, runbooks, golden cases. The harness is what you keep. The horse is what you rent. After a kill switch you find out which one you actually built.

Notice what is missing: chat logs in a vendor workspace. Fine-tunes you cannot export. "Custom GPTs" that were never files. System prompts living in a CoE Notion. Those felt like assets in May. In June they were folklore.

Skills were the surprise survivor

I spent the first half of 2026 arguing that you should build skills, not agent armies. The recall made the argument rude. A skill is a directory. Directories do not care which lab answered the letter. The agent that could still read SKILL.md on Monday, Claude Code, Cursor, a local runtime, a thinner open-weight loop, kept the procedural knowledge. The agent that "knew how we work" only because it was the approved chat did not.

If you did nothing else after Fable, export the work into folders. Name, description, scripts that score, references the model loads when relevant. That is not tidiness. That is workflow ownership on disk. We will need that sentence again when the same physics hits a plant that cannot call the API.

Europe bought the nerfed tier. The trench did not wait.

The second-order forecast from the Fable essay landed on schedule. Allied markets purchased the filtered SKU and slept. Practitioners pulled portable weights into the same pipelines they already used for classification, summarisation, and internal tools. Press releases called the filtered SKU a milestone. Operators called it Tuesday.

You can run both. You should. Market A for the invoice and the audit trail. Market B for the path that still works when Market A gets a letter at 5:21pm. Treating those as a culture war is how architecture decks avoid the word discontinuity.

The failure modes (you already have them)

Freeze as governance. Pause all tools. Shadow IT rises. You will spend Q3 discovering what your people already did without you.

Vendor as architecture. "We standardised on X." You standardised on a permissioned half-life. Dual-home the route or admit you are renting.

Eval as theatre. A bake-off slide from March. No golden cases from last week's actual tickets. When the model changes, you cannot tell if you got worse.

Capability in the chat transcript. The only copy of how the brand check works is a thread. That is not a system. That is a campfire.

What to do before the next letter

You do not need a two-year sovereignty programme. You need an inventory this month:

• List every production-shaped loop. Name the model route. Name the fallback.
• Move procedural knowledge into skill folders the agent loads, not into a vendor workspace.
• Keep a small, honest eval suite per loop. Run it when the SKU changes, not when a committee asks.
• Decide now which work is allowed to run on open weights, on-prem, or a degraded frontier tier. Write it down. Waiting for the next directive is how you donate the decision to panic.

Then look past the office. The next discontinuity is not another chat SKU. It is the site that cannot send a drawing to Virginia, the cell that is air-gapped, the qualification lot that does not care about your Copilot agreement. Software just learned the lesson. Hardware is next.

The kill switch taught a simple accounting identity. Rent the horse. Own the harness. If it is not on disk, evals, routes, weights, skills, it was never yours. The letter was loud. The folders were quiet. Quiet is what compounds.

Related: The Fable Directive, Nerfed Frontier, Open Weights, and Skills Became Infrastructure.