Your laptop had to be awake

Through the winter Anthropic built out Cowork, the no-code sibling to its Claude Code tool, into an enterprise system: an admin-controlled marketplace where IT can push a finance agent to only the finance team, a dozen new connectors into Gmail, Calendar, DocuSign and the rest, and recurring scheduled tasks — an agent that pulls your overnight Slack and email into a report before you wake up.

In the first build the agent stopped for a manual approval prompt on every single action — and a laptop-closed cloud version only arrived about seven weeks later.

The pitch was an employee that works while you sleep. The first version did not. Scheduled tasks ran only while the user's own computer was awake and the desktop app was still open; if the laptop slept, the 6 a.m. briefing silently skipped. A worker testing it hands-on hit permission errors, server failures, and a manual approval prompt for every single action. The always-on agent needed you to leave your machine on all night.

Anthropic seemed to agree it had shipped the wrong thing. Roughly seven weeks later it released a separate cloud version that actually runs with the laptop closed — the product the scheduling feature was sold as from the start.

The gap is the story of enterprise agents right now. The marketplace and the connectors are real and useful, and governed per-team provisioning is what cautious IT departments actually want. But the distance between the demo and the first build is wide enough that Anthropic's own head of Americas conceded the 2025 promise that agents would transform the enterprise turned out to be 'mostly premature.' The capability is arriving; the autonomy in the marketing was running about two months ahead of the code.