AI assistant outage, project blocked: August 16 and what stays with you

On August 16, 2026, Claude Code was unavailable for forty-two minutes. For anyone having software built by an agent team, the question that follows matters more than the incident: depending on where your documents and code live, a provider outage takes away an evening or your whole project.

AI assistant outage, project blocked: on August 16, 2026 at 21:58 UTC, Anthropic confirmed an incident affecting Claude.ai, Claude Code, and Claude Cowork, with degradation reported at 22:07 and full recovery at 22:40 (BleepingComputer, August 16, 2026). Forty-two minutes in which nobody had a line written.

What happened on August 16

The incident began at 21:58 UTC. Nine minutes later, the publisher reported authentication problems followed by degraded performance; full recovery came at 22:40. The administrative console and key-based access remained operational according to the same account, without the article explaining why. Forty-two minutes on an August evening do little damage. The same interruption on a delivery Tuesday, or three hours instead of forty minutes, raises a different question, not about provider reliability, which nobody controls, but about what remains within reach when it goes down.

The next day, GitHub went down for seven hours

On August 17, 2026, GitHub was unavailable for 7 hours and 47 minutes: website, authentication, APIs, automated jobs, issues, and Copilot. The publisher released its analysis three days later and ruled out code or configuration changes: the cause was capacity, under a load that saw monthly contributions rise from 1.4 billion in April to 2.9 billion. Two outages in two days at independent providers, affecting the two components on which much of the world's software work depends. A business owner has no control over these incidents. They do control where their files live.

What you lose depends on where your project lives

Compare two situations. You build on an online platform hosting the editor, code, database, and public address: when the provider fails, you cannot work, reread yesterday's writing, or show the result to a partner, and your live application goes down with it. You work in an application installed on your Mac: the assistant falls silent, so no more code is written, but the approved specification, step breakdown, and version history remain readable and editable on your disk. In the first case the outage takes away the worksite. In the second it takes away the worker, and you can spend the evening reviewing what you were about to have built.

What Maestro cannot do either

No tool produces text without a model. Maestro has no engine of its own: it controls Claude Code, Codex, Cursor, or Mistral Vibe, and when the chosen provider goes down, construction stops too. A hosted tool even has a real advantage here. Your online application continues serving customers if the outage only affects the editor, whereas a project on your Mac waits for someone to deploy it somewhere, a step that remains your responsibility. The difference concerns the material you hold, not the availability of the engine that produced it.

Change assistants without changing projects

Maestro lets you choose among four assistants and switch mid-conversation, with a subscription or a key stored in macOS Keychain (your projects do not leave your machine). On the evening of August 16, the move was to switch providers and resume the current step, since the brief, specification, and breakdown depend on none of them: they are French documents, not model-specific settings. This is the reasoning we used for a tool that loses its engine overnight. The choice depends less on today's model quality than on how easily you can switch without losing anything.

The one thing to do during an outage

Open your project folder and see what is there without a connection: brief, specification, step breakdown, code, version history. Whatever you can read offline is what you will retain during a long outage, and the list takes ten minutes. If it is empty because everything lives with a publisher, ask them in writing today how to export your work, and keep the answer with your contracts, as we described in what happens to an application when the tool disappears. A forty-minute outage costs an evening: it is the cheapest moment to discover what you could not retrieve.