Detailed Analysis
Anthropic's Claude Cowork feature experienced a partial outage on April 16, 2026, preventing the service from starting for a subset of users. The incident was first reported at 19:24 UTC, with Anthropic's engineering team identifying the issue and moving toward a fix by 20:47 UTC — roughly 83 minutes after initial detection. The outage was isolated to Claude Cowork running on Claude Desktop version 1.3036.0, while core services including Claude.ai, the Claude API, and Claude Code remained fully operational throughout the event. No estimated resolution time was publicly communicated, though the official status page at status.claude.com confirmed active remediation efforts were underway.
The error patterns reported by affected users cluster around virtual machine connectivity failures — most prominently messages such as "Failed to start Claude's workspace — Can't reach the Claude API from Claude's workspace" and "sdk-daemon not connected." These errors point to the architectural underpinning of Claude Cowork, which relies on a local virtualization layer (the CoworkVMService) to spin up an isolated compute environment on the user's machine. On Windows in particular, this service has exhibited fragility, including a known tendency to stop after system sleep cycles, crashes, or automatic updates. User-reported workarounds — such as manually restarting CoworkVMService via Windows Services and verifying that Hyper-V is enabled — provide temporary relief but underscore the complexity of managing a VM-dependent feature across diverse consumer hardware configurations.
The broader context of this outage reflects the inherent challenges of shipping a product like Claude Cowork, which remains in research preview on Windows as of early 2026. With over 60 open GitHub issues and documented regressions tied to recent auto-updates (including the v1.1.5749 March 2026 release), Claude Cowork occupies an early, unstable phase of its lifecycle. The feature represents Anthropic's ambition to give Claude persistent, sandboxed compute — enabling it to run code, manage files, and execute long-horizon tasks autonomously — but the infrastructure demands of local virtualization introduce failure surfaces that a purely API-based product would not face. Mac support is further restricted to Apple Silicon (M1 and later), leaving Intel Mac users unsupported entirely.
This incident fits into a wider industry pattern in which AI companies are racing to move their assistants beyond stateless chat interactions toward agentic, environment-aware deployments. Features like Claude Cowork, OpenAI's local code execution tools, and similar offerings from Google require tight integration between cloud services and local system resources — a stack that is inherently more brittle than browser-based SaaS. As Anthropic scales Cowork from preview to general availability, hardening the virtualization layer, improving service auto-recovery, and expanding platform compatibility will be critical prerequisites. The April 16 outage, while partial and relatively brief, illustrates that the reliability bar for agentic compute infrastructure must rise substantially before such features can be trusted for professional or mission-critical workflows.
Read original article →