A workflow that loses task state when it loses model access is poorly prepared to continue working. On June 12, 2026, Anthropic announced a global suspension of Fable 5 and Mythos 5 to comply with an export directive received that day. Other models remained available. For anyone depending on that access, the plan changed immediately.
Five days later, on June 17, 2026, I asked for agents stopped by Claude usage limits to resume. That was a usage limit, a different cause from the announced suspension. The situation matters to me because the next request was to continue the work. Nobody wants a provider interruption to become an extra archaeology session to discover where the task stopped.
When an agent stops, I want to start with verifiable state. Reading material creates one kind of resumption cost. Having already changed something creates another. A message saying it was nearly finished helps little without a way to check what exists. Restarting blindly can duplicate effort; continuing blindly can carry forward a conclusion that was never verified.
Work that would be expensive to reconstruct should therefore leave a record understandable outside the original conversation. The objective belongs there, alongside what has been checked and the uncertainty preventing the next step. That record should support continued investigation. Copying the first executor's confidence does not transfer the evidence it should have produced.
An alternative also requires honesty about its conditions. Another model may need different review or a smaller assignment. Hiding the switch to preserve a feeling of continuity makes it harder to understand why delivery changed. I prefer to know what changed in the executor and adjust the work explicitly.
The record should fit the task. I will not build an enormous structure for a short edit that is easy to check directly. A long investigation deserves enough preserved information to avoid depending on one window's memory. The reference point is the cost of recovering the work, with the remaining verification still visible to the next executor.
The June suspension and my later request to resume work have separate causes. Both lead me to the same operational requirement: before distributing a long task, I want state that can survive an interruption. On resumption, the next agent starts by checking that state and proceeds from the next actual uncertainty. My favorite model can become unavailable; the task still needs to remain understandable.