Computer-use agents: magic in demos, but what actually survives daily use?
Grand-Produce-8527 · reddit · 2026-09-09
After months of experimenting with Claude computer use and browser agents, a developer finds a persistent gap between demos and reality: logins trigger bot detection and CAPTCHAs, public data is faster via curl/Playwright, high-stakes sites aren't worth unsupervised LLM clicks, and multi-step no-API tasks tend to break around step 6. Candidate real use cases proposed: internal admin panels without APIs, repetitive vendor form filling, smoke-test QA, change-watching pages with no feed, personal admin (insurance claims, DMV, refunds), and legacy desktop apps. They also ask whether a persistent real browser profile is the difference between working and being instantly blocked.
More from coding & agent
- Microsoft's MCP Live! half-day event covers building servers, auth and enterprise readiness — lee_stott · 2026-09-09
- NVIDIA unveils CUDA Rust: cuda-oxide and cutile-rs, two tracks for writing GPU kernels — ssh4net · 2026-09-09
- Agent pauses should surface the decision, not just 'I need clarification' — OriginalHospital · 2026-09-09
- Self-hosted podcast search stack: Whisper + pyannote + pgvector indexes 180 hours — Apprehensive-Ear9499 · 2026-09-09
- Dolphin creator Eric Hartford unveils With, a systems language compiling to native code — QuixiAI · 2026-09-09
- TokEMS: Open-source self-hosted conference platform with 140 iterations, from registration to check-in — vista8 · 2026-09-09