On April 15, Claude stopped working. Not a piece of it — all of it. Claude.ai, the API, Claude Code, Desktop, Mobile — every surface went dark from 10:53 AM to 1:42 PM ET. Over 5,100 Downdetector reports at peak. Managed Agents running customer workflows? Dead. Routines processing overnight tasks? Dead. Your team's IDE? Also dead. Same moment, same root cause, same vendor.

Most outages earn a shrug and a coffee break. This one earns a harder look — because of what happened in the two weeks before it.

The Stack That Ate Your Plan B

Back in February, a Pragmatic Engineer survey of 15,000 developers named Claude Code "most loved" AI coding tool at 46%, lapping Cursor (19%) and GitHub Copilot (9%). Nobody seriously debates whether the tool works. The question is what happens when it doesn't.

Count the layers: Claude Code, Claude Desktop, Claude Routines, Claude Managed Agents. Every one runs on one model family, one vendor, one API. If you adopted Routines on April 14 — exactly one day before the outage — you added another dependency to the same single point of failure.

The Two-Week Fork

Between April 2 and April 15, the industry split.

Anthropic went deeper into vertical integration — building every layer of the stack and locking them together. On April 8, they launched Managed Agents: cloud-hosted AI workers at $0.08/hour plus token costs. Notion, Rakuten, Sentry, Asana, and Atlassian signed up day one. On April 14, they shipped Routines and a redesigned Desktop — saved configurations that run on Anthropic's cloud while your laptop sleeps. Pro gets 5 per day, Max gets 15.

Everyone else went the opposite direction. On April 2, Cursor 3 shipped with GPT-5.4, Claude Opus 4.6, Gemini 3 Pro, and Grok Code — an "Auto" mode picks the best model per prompt. On April 15, hours after Claude came back online, OpenAI updated its Agents SDK to support 100+ LLMs (large language models — the brains behind ChatGPT, Claude, Gemini) via LiteLLM, with seven sandbox providers. Zed Pro serves six models for $20/month.

One camp bet on depth. The other bet on optionality. Then the deep camp's single dependency went offline for three hours.

The Vertical Bargain

Claude Code is better because it's locked in. That's the deal, and you took it. When Anthropic optimizes prompt chains, caching strategies, and tool calling for exactly one model family, they skip the translation layers and lowest-common-denominator compromises that multi-model routing demands. One model, tuned end to end.

As Nate's Newsletter observed on April 8: "behavioral context creates switching costs deeper than anything Microsoft or Salesforce ever built." Claude Code doesn't just lock in your subscription — it locks in your muscle memory, your CLAUDE.md configs, your Routines, your team's entire workflow.

Optimization is concentration. And on April 15, the bill came due.

128 Reasons to Worry

The three-hour outage wasn't a freak event. Anthropic's status page logs 128 incidents in the last 90 days — 41 major outages, 87 minor — with a median duration of 1 hour 7 minutes. That's roughly one event every 17 hours.

VentureBeat warned on the day Managed Agents launched that enterprises adopting the platform "lose control, observability and portability and risk further vendor lock-in."

Nobody listened. The product was too good.

But now you have Managed Agents handling customer workflows, Routines running overnight jobs, and an entire team whose coding muscle memory lives inside Claude Code. Each new Anthropic product you adopt widens the blast radius of the next outage — and narrows your options for surviving it.

The Audit You're Avoiding

Before your next sprint, count the workflows that depend on Claude with zero fallback. Your IDE. Your scheduled tasks. Your agent fleet. Your Slack and Notion integrations through Managed Agents.

The best developer experience in AI is also the biggest single-vendor bet in enterprise software. In the same two-week window, the rest of the industry voted for optionality — multi-model, multi-provider, swap-on-failure. That approach adds complexity and degrades quality. But it has a Plan B.

On April 15, for two hours and forty-nine minutes, you found out what it's like to not have one.