Show HN: Agent.exe, a cross-platform app to let 3.5 Sonnet control your machine

Original Article ↗ Hacker News Discussion ↗

A new cross‑platform app, Agent.exe, lets Anthropic’s Claude 3.5 Sonnet remotely control a user’s computer via screenshots and simulated input, enabling tasks like booking flights or online shopping. Commenters are split between excitement over automation and accessibility potential, and alarm at the security and privacy implications of effectively installing “consensual malware” with access to browsers, bank accounts, and personal data. Early testers report the system is slow, error‑prone, and sometimes confidently wrong, raising questions about reliability, cost, and how such agents should be sandboxed in practice.

Overview & Concept

  • App exposes Anthropic’s new “Computer Use” capability via an Electron desktop client, letting Claude 3.5 Sonnet operate your OS (open browser, click, type, etc.).
  • Intended uses discussed: flight and gift shopping, Amazon cart filling, monitoring sites (e.g., CD rates, pandemic-era Amazon slots), UI testing, and general “assistant” work.

Setup, UX & Platform Issues

  • Some disappointed setup isn’t as simple as running a single binary.
  • Coordinate mapping is buggy on macOS, Windows, and in sandbox/VM; clicks often miss targets, especially on non-standard resolutions.
  • The app’s own window frequently obscures parts of the UI, causing the model to misread state; suggestions include auto-hiding the window, capturing only a target window, or using a less intrusive frame/toolbar.

Security, Privacy & Banking

  • Many call it “malware-like” or a “botnet waiting to happen,” warning not to run on a primary machine.
  • Recommended mitigations: separate user account, no sudo, dedicated VM, and possibly network isolation.
  • Debate over banking risk: EU posters cite strong 2FA (PSD2, hardware tokens, apps); US posters note “trusted device” flows and weak per-transaction 2FA.
  • Concerns that LLMs can easily leak sensitive data while “doing useful work,” with no AV/firewall model for this risk.

Behavior, Reliability & Safety Rails

  • Claude shows quirky “personality”: preference for Firefox, occasional wandering into Yellowstone photos.
  • Safety rails block some actions (e.g., sending Discord/WhatsApp messages), but users question the rationale and completeness.
  • Reliability is poor: wrong flight dates, mis-clicks in CAD/3D tools, confusion between search and message fields, inability to verify results, yet still declaring success.
  • Works for trivial tasks (reading system time), fails surprisingly often on slightly more complex ones.

Cost & Performance

  • Reported costs: ~$0.38–$0.50 for simple multi-step tasks; latency is several seconds per action.
  • Some liken current costs to a high hourly rate for an unreliable assistant; others expect prices and quality to improve, though this is disputed.

Broader Implications & Ethics

  • Split between excitement (“start of Skynet,” major productivity and testing potential, big shift in software work) and alarm (Pandora’s box, intentional malware, undefined/non-deterministic behavior normalized).
  • Several see strong accessibility potential (hands-free computer use, voice+LLM hybrids), while others argue ethics and safety training for developers are lagging far behind these capabilities.