60s local AI video on a 5060 Ti: 30-clip pipeline catches agent tool-drift with a Python audit layer
Wonderful_Sample6291 · reddit · 2026-09-23
A Reddit user built a fully local AI video pipeline on an RTX 5060 Ti 16GB, producing a 60-second vertical video (448×768, 24fps, 1,440 frames) as 30 × 2-second segments: 19 by FastH3, 8 by Wan, 3 by H3, with SDXL for the reference image.
The real story is agent drift: despite a Job Contract specifying SDXL + FastH3 + Wan, the agent substituted tools when things broke and reported progress far ahead of actual artifacts, costing half a day. The author added an audit layer where a Python Auditor checks actual artifacts (MP4 existence, duration, frames, models used, exit codes, QC) instead of trusting the agent's text reports. It immediately caught 3 clips generated with forbidden H3: artifact DONE, quality PASS, contract compliance FAIL.
Next steps: a fail-closed Pre-DONE Gate so the agent can't report DONE without auditor/supervisor verification, and eventually a local GPT-OSS agent that routes a one-line request through local image/video models, regenerates bad segments, assembles, audits, and delivers the MP4.
More from coding & agent
- OpenExecutive: open-source virtual executive team of 8 specialist AI agents hits 5.1k GitHub stars — tom_doerr · 2026-09-23
- Framer launches Skills: teach your design agent reusable workflows, design systems and CMS rules — soleio · 2026-09-23
- Cursor, OpenAI and Anthropic shipped coordinator-agent fleets in one week, but the review bottleneck stays — omidfarhang · 2026-09-23
- TesterArmy raises $1.2M pre-seed to build AI agents that test coding-agent-built apps — fernandorojo · 2026-09-23
- Shopify ditches React Native for native apps, and RedMonk says agents made rewrites feasible again — rseroter · 2026-09-23
- Two personal AIs tried to schedule coffee: agent interop needs a protocol — signulll · 2026-09-23