Detailed Analysis
A user running Claude's Cowork feature has surfaced a friction point in Anthropic's approval system: scheduled task creation appears to bypass the "skip all approvals" setting, forcing manual confirmation even when the user has explicitly configured Claude to operate autonomously. The reporter describes a workflow where Cowork runs recurring scheduled tasks and has been instructed, via custom prompt, to spin up additional one-off follow-up tasks as needed. In practice, however, each new task still triggers a confirmation dialog requiring a human click — which breaks the entire premise of unattended automation. If a task is meant to run and self-schedule follow-ups while no one is watching, a blocking approval prompt means the follow-up simply never gets created.
This is a meaningful gap because it exposes an inconsistency in how granular Anthropic's permission controls are across its product surface. The user draws a direct comparison to Claude Code, which already supports allowlisting specific tools or actions so that trusted operations skip confirmation prompts entirely while riskier ones still require sign-off. That precedent shows Anthropic has already built the underlying philosophy — treat "skip all approvals" as a blunt instrument and let power users fine-tune which actions are trusted. Cowork's scheduled-task creation flow apparently hasn't inherited that same nuance, suggesting the feature may have been built with a blanket assumption that task creation is inherently sensitive enough to always warrant human review, regardless of what the global approval setting says.
The underlying tension here is a recurring one in agentic AI products: the more autonomy a system is given, the more its safety rails can become the very thing that undermines its usefulness. Anthropic has been positioning Claude increasingly as an agent capable of multi-step, long-horizon work — scheduled tasks, computer use, and now Cowork's more persistent, self-directed workflows — but each new autonomous capability raises the stakes on getting permission granularity right. A confirmation dialog that can't be suppressed doesn't just annoy users; it silently breaks the automation loop, since the whole point of scheduling a follow-up task is that no one is present to approve it. For asynchronous or overnight workflows, this could mean tasks quietly fail to propagate without any visible error.
More broadly, this complaint reflects a common growing pain among AI products racing to add "agentic" features: shipping the capability to act (create tasks, call tools, take actions) faster than shipping the corresponding controls to govern that capability safely and flexibly. Users who push these systems hard — running multiple scheduled tasks daily, chaining follow-ups programmatically — are the ones most likely to hit these seams first, and their feedback (as seen in this Reddit thread) often serves as a leading indicator of where product teams need to invest next. Anthropic's own Claude Code already demonstrates the target architecture: scoped, tool-level allowlisting rather than an all-or-nothing toggle. The expectation from power users is that Cowork will eventually converge on that same model, either by expanding "skip all approvals" to cover task creation or by introducing a dedicated permission specifically for autonomous scheduling, closing a gap that currently limits how "hands-off" the product can truly be.
Read original article →