giving an AI agent a key to your wallet — spoolcast dev-log #8
devlog about letting AI agents spend real money in PipelineCPC, and the fences that keep it from going wrong.
PipelineCPC needed to let AI agents research keywords on a user's behalf, but every lookup spends real money. The session was not about building the integration; the endpoints already existed. The real work was building fences.
This devlog walks through what those fences look like in practice. Naming what an agent must never be able to do. Defaulting every control to fail closed. Three-tier credit caps. One-time hashed secrets. A 5-key cap. A Codex review of the threat model. And the bugs that only appear when a real person uses the feature instead of a test.
The Spoolcast pipeline is itself agent-driven, so the episode connects back to that — the same kind of agent doing the production now needs the same kind of fences.
AI disclosure
This video was made with Spoolcast, the AI video system I am building. I did not manually edit it in a traditional timeline; the system helped produce the structure, script, visuals, narration, render, thumbnail, and shipping files from source material, direction, feedback, and approvals.
How it's made
Built on the Spoolcast illustration/chunk Remotion pipeline with generated scene images, Google Chirp3-HD chunk-level SSML narration, rendered widescreen output, narration-only SRT captions, and a final YouTube thumbnail.