← back
How to avoid disaster when vibe-coding a billing engine — Andrew Garvin, Stripe
Takeaway
Use coding agents to accelerate realistic billing experiments, supported by skills, clear errors, and sandbox validation.
Summary
- Stripe Projects provisions services through a CLI, with a demo configuring Metronome billing from natural-language instructions.
- The example recreates a Lovable-style prepaid credit and automatic-recharge model in a sandbox.
- Portable skill files guide agents through a complex billing API, while descriptive errors help them recover from setup failures.
- Generated test usage makes billing behavior observable; business-critical logic still requires human review before production.
usage-billingstripemetronome
Original description
Andrew Garvin types one sentence asking for a billing engine that copies Lovable's pricing, and gets back a working sandbox: a customer, metered usage flowing in, and a draft invoice broken into separately scoped credit pools for builds, plan mode, cloud and gateway calls. Reproducing that by hand means understanding auto recharge, credit expiry and overage, which is exactly where people hurt themselves. He cofounded Metronome, the usage billing platform Stripe acquired this year in its largest deal, so he has watched a lot of companies get this wrong. What makes the talk useful is that a billing vendor stands on stage and argues against full autonomy. Billing carries deep business logic and real money, so his recommendation is to let an agent accelerate you into a test environment and stop there, rather than ship to production unattended. The guardrails are unglamorous: portable skills files that carry the API's hard won context, and deliberately verbose error messages written so an agent can correct itself. He also separates three things people blur together, an agent as your product, as your buyer, and as your user. The third is the disruptive one, and he points at a large software company cutting seat prices and moving to credits, because seats stop meaning much once one agent does the work of many logins. Speaker info: / agarvin Timestamps: 0:00 - Metronome, Stripe, and what they are demoing 2:29 - Where billing goes sideways with agents 3:40 - Provisioning an environment from the CLI 6:02 - Skills files that carry the hard parts 7:13 - Why a human stays in the loop 9:30 - Agent as product, buyer and user 10:38 - Moving off seats and onto credits 13:01 - Opening what the agent actually built 14:12 - Scoped credit pools and a draft invoice 16:27 - What this suggests beyond billing