Andrew Garvin

Andrew Garvin is a co-founder of Metronome, the usage-based billing company that became part of Stripe. His work connects software consumption with the commercial decisions behind it: how to meter usage, apply credits and commitments, and ensure that bills reflect the intended pricing model. He brings a finance operator’s perspective to agent development, where automated systems can both generate costs and change how customers use software.

From finance to billing infrastructure

Garvin’s career moved between investment firms and startup operations. In his Venture with Grace interview, he describes his first job out of college as working at Peter Thiel’s hedge fund. The episode’s host-provided description lists roles at Clarium, director of finance at Halcyon Molecular, CFO at Mithril Capital, and work as a fractional CFO and adviser, without establishing their subsequent order. These roles span financial management in operating companies and investment businesses.

His work advising Blend brought the operational difficulties of consumption-based pricing into focus. Sales compensation and financial planning depended on more than issuing invoices: teams needed to measure consumption, manage pricing changes, explain them to customers, and compensate salespeople whose accounts generated variable revenue. Those problems helped shape his approach to building Metronome.

Garvin and fellow co-founder Scott Woody knew each other from Berkeley. Their discussions about starting a business converged on billing, combining Woody’s experience with growth at Dropbox and Garvin’s experience with consumption-based businesses. Garvin’s early responsibilities at Metronome included determining its first sales hire and its own pricing model, before expanding across operations and revenue.

Metronome’s work became particularly relevant to AI services, where customer activity can produce substantial and variable computing costs. The company meters API calls for OpenAI and Anthropic and supports credits, commitments, discounts, and other commercial terms. Stripe completed its acquisition of Metronome in January 2026. Garvin’s subsequent work with the Stripe team explored how coding agents could make this complex infrastructure easier to configure and test.

Pricing that can evolve with the product

Garvin argues that continuous pricing evolution requires product development, commercial experimentation, and financial accuracy to stay aligned. His position on hybrid pricing extends beyond choosing between subscriptions and consumption charges. Seat-based credits combine access with an allowance of usage; account hierarchies organize how a pricing model applies across related accounts; packages bring commercial components together into an offer. These mechanisms give teams ways to change what they sell and how they charge as products develop. For Garvin, billing infrastructure must support that experimentation while preserving an accurate account of what customers owe.

That concern becomes sharper when agents change the relationship between software access, activity, and value. Garvin distinguishes agents as products, buyers, and users because each role creates a different business problem. An agent sold as a product can run up a token bill, making metering and spending controls important. An agent acting as a buyer can procure backend services, so providers need to make those services discoverable and possible to provision automatically. An agent using an existing application can concentrate work previously performed by many people into one automated user, weakening the relationship between seat count and value delivered.

He uses Metronome’s work with HubSpot to illustrate that last shift, describing its movement toward credit-based pricing as agents increasingly operate software on users’ behalf. The underlying question is how a software company charges for work when the number of people logging in no longer captures how much the system does. Prepaid and postpaid commitments for enterprise AI customers extend this pricing discussion beyond self-service credit purchases.

Making agent-assisted billing inspectable

Garvin’s Stripe Projects demonstration made agent-assisted billing concrete. A natural-language request asked a Metronome billing agent to create a sandbox modeled on Lovable’s pricing. The agent configured a test customer, credits scoped to different kinds of consumption, simulated usage, and a draft invoice. Usage drew down the credit balance, allowing a person to inspect the resulting charges.

His approach rests on several practical requirements:

• Testing with realistic usage. A plausible configuration alone cannot show whether credits and invoices behave as intended. Populating the sandbox with usage lets a team inspect how consumption becomes charges, then adjust the model before exposing real customers to it.

• Portable skills files. Reusable instructions supply configuration guidance and direct the agent to populate test usage. Clear, informative errors help it recognize failures and correct its requests. Together, these mechanisms reduce the work needed to establish a useful test environment and help an agent recover from mistakes around a complex billing API.

• Human review before production. Billing contains business-critical logic. Garvin’s demonstration created a sandbox for inspection and adjustment; deployment remained a human decision. He also identifies runaway agent spending as a growing concern and proposes bounded wallets as a possible control.

Garvin’s work joins faster implementation with financial oversight. An agent can help construct a pricing model and exercise it, but teams still need to understand the resulting charges, judge whether they match the business they intend to run, and decide when the configuration is ready for real customers.