Local uncensored Qwen writes the pitch, 8 cloud Opus workers fill 107 directory forms in 2 hours

GapNew4766 · reddit · 2026-10-17

A developer shares a multi-agent setup that completed 107 directory submission forms: one JSON call fans out to 8 isolated worker processes, each driving its own headless Chrome instance with its own UA string, solving image CAPTCHAs via a vision model, and writing results to disk.

The two-model split (the core idea)

Orchestrator (local): uncensored Qwen3.8-27B, Q4KM GGUF on an M5 Pro with 36GB unified memory, 28 tok/s, $0 API cost. Reads directories.json, plans waves of 8, writes each task instruction verbatim — including the exact marketing copy to type — and picks category and UA rotation
Workers (cloud): 8 Claude Opus 5 instances handling mechanical reliability — opening Chrome, typing exactly what they were told, waiting for DOM load, retrying on timeout
Why the split? Cloud models hedge and moralize when asked to write the pitch themselves; the local uncensored model writes copy like "It listens to your meetings and writes notes so you stop pretending you were paying attention," then bakes that exact string into the worker's instruction

Worker isolation: each worker sees only its own instruction string; they coordinate via files on disk (./out/result{id}.json).

CAPTCHA handling: image CAPTCHAs (hCaptcha, FunCaptcha) are screenshotted, read by a local vision model, and typed in, with up to 3 retries; ReCAPTCHA v2 checkboxes remain unreliable — the model can read the image but can't reliably click "I'm not a robot," so failures are logged as captchafailed and skipped.

A useful pattern worth stealing: a cheap local model for planning and content generation, expensive cloud models for reliable execution.

Original post →

Local uncensored Qwen writes the pitch, 8 cloud Opus workers fill 107 directory forms in 2 hours

More from coding & agent