My agents burned video credits on bad frames, so now they must ask before spending
coolxeo · reddit · 2026-10-01
The author runs a small multi-agent pipeline producing toddler videos: one agent plans, one generates stills, one checks them, one renders video. Early on, agents auto-triggered the next step, and wrong-character stills kept flowing into the most expensive step — video rendering.
The fix: keep cheap steps (planning, still checks) fully automated, but require human approval for anything that spends real credits (rendering, final stitching). The agent posts a card in shared chat stating what it wants to run, why, and which checks passed or failed; nothing runs without a tap.
Most rejects now land at the still stage before any expensive compute runs, costing only a few seconds per approval. He asks where others draw the automation boundary: on the plan, on QA, or right before publish.
More from coding & agent
- Why Shopify dropped React Native: AI agents made native the future again — davemccollough · 2026-10-01
- ClawDaddy MCP lets AI agents buy domains via Stripe and manage DNS end-to-end — modelcontextprotocol · 2026-10-01
- MetricDuck MCP serves SEC financial data to AI agents with filing-level traceability — modelcontextprotocol · 2026-10-01
- A 1-cent verifier catches 61% of AI agents falsely claiming task completion, paper finds — alex_verem · 2026-10-01
- Giving Claude full PC control built a playable SCP-096 horror game in 3 hours — imjustnewatai · 2026-10-01
- OpenAI dots experiment: a handoff-note template to test agent memory across sessions — Kitchen-Jicama8715 · 2026-10-01