Detailed Analysis
A Reddit user's complaint about Claude in Chrome's sudden behavioral change toward Facebook Marketplace automation has surfaced a meaningful inflection point in how Anthropic is managing its browser extension product. The user describes a coworker automation workflow — using Claude in Chrome to post listings on Facebook Marketplace — that functioned smoothly for over a month before abruptly requiring manual confirmation before every single action. The change was not announced through any visible update log or documentation, leaving users to piece together the cause through independent research. The user's conclusion, corroborated by broader community discussion, is that Anthropic has implemented or tightened restrictions on automating certain platforms, with Facebook specifically appearing to be among those blocked or throttled for security-related reasons.
The timing of this disruption aligns closely with a significant expansion of Claude in Chrome's capabilities. The extension recently exited an extended testing phase and entered beta availability for all paid subscribers — Pro, Max, Team, and Enterprise tiers — introducing features like workflow recording, scheduling, multi-tab support, Claude Code integration, and a planning mode that allows independent execution after an initial approval step. This planning mode, which requires user sign-off before autonomous action chains begin, is almost certainly the mechanism behind the confirmation prompts the user is now encountering. What was previously a more permissive testing environment appears to have been standardized under a more cautious interaction model as the product moved toward formal beta release — a classic pattern in AI product maturation where autonomy is incrementally gated behind human-in-the-loop checkpoints.
The friction the user describes reflects a fundamental tension Anthropic is navigating between empowering automation and managing liability, platform compliance, and reputational risk. Facebook Marketplace, like many major platforms, actively prohibits automated posting via its terms of service, and Anthropic's decision to restrict or flag such workflows is likely as much about legal and partnership exposure as it is about user security. The user's frustration — that risk tolerance should be the individual's choice — is a common refrain in the power-user community, but it undersells the structural reality: Anthropic is building a product intended for broad enterprise deployment, where unchecked automation against third-party platforms creates compounding legal and trust risks that no individual user's consent can fully absorb.
The broader trend here is one of increasing deliberateness in how frontier AI labs ship agentic tools. Anthropic's iterative approach — recording, planning mode, confirmation prompts, CAPTCHA pauses — reflects a philosophy of keeping humans meaningfully in the loop during the current phase of agentic AI deployment. This stands in contrast to more permissive automation frameworks like Playwright or Puppeteer, which the user considers as alternatives. Those tools offer raw browser control without AI-layer restrictions, but they carry their own risks: platform detection, blocking, and the absence of any intelligent error recovery. The user's dilemma — a polished but restricted AI tool versus a flexible but detectable lower-level tool — illustrates precisely the gap the broader agentic AI ecosystem is still working to close, where neither fully autonomous AI agents nor traditional scripting cleanly serves the nuanced, adaptive needs of real-world workflow automation.
Read original article →