Kill the Finish-Line Price. A Relay Has a Definition Per Handoff.

David Haber at a16z wrote this morning about long-running agents winning what he calls the relay race: customer journeys that run for weeks or months, where value leaks at every handoff and only shows up at the end. The product argument is strong. The pricing conclusion at the end of it is the one I'd put on the kill list.

The short version

David Haber's a16z piece "Winning the Relay Race" argues that long-running agents can own months-long customer journeys end to end and that, since the outcome is observable and attributable, vendors should be paid on it; his clean unit is a settled plaintiff case, whose value he calls "clearly identifiable." James Brown's Stripe write-up of Monetize 2026, published the day before, reports what happened to the cleanest single-outcome meter in the market: per Aisling O'Reilly, Fin billed per resolved conversation until the agent did more than resolve, then split into multiple billable outcomes and a credit pool. I priced one outcome, a resolved ticket with a 48-hour definition. It held for 98% of cases, and the other 2% became more than 200 contested cases a week and a dispute team that grew from 4 to 14. A relay has a definition at every handoff. Kill the finish-line price. Keep the finish line as the renewal number and price the leg whose definition the customer will sign.

Haber's race

The piece is worth reading for the product half. The relay is a journey with many steps and many hands: in plaintiff law, intake specialists gather facts, an attorney decides whether to take the case, paralegals order police reports and chase witnesses, treatment has to conclude before the medical chronology, and the demand letter comes last. Value leaks at each handoff. The firm that returns the call fastest signs the client, nobody makes the fifth call to a witness, and a routine case gets settled at the first offer because nobody read the doctor's note that pointed at a brain injury.

A long-running agent, in his telling, is assigned to the case at intake with the goal of maximizing recovery, wakes on events and deadlines, dispatches sub-agents to call witnesses and update the client, keeps the full record, and learns across cases.

Then the economics. "As impact is observable and attributable to agents, vendors can participate in the upside they create by pricing based on work products or outcomes. In plaintiff law, for instance, a single case is a discrete unit of work and the value of the resulting outcome, the settlement amount, is clearly identifiable."

He's careful after that. Most states bar non-lawyers from sharing in fees, so pricing would have to align with settlement values without taking a share. Procurement has no precedent. "Attribution debates will likely surface initially." And he lays out three routes to market: sell the platform, sell the service, or go full stack and buy the debt portfolio yourself.

Identifiable is the word I'd stop on. Identifiable isn't agreed. I've written that line before about a single ticket, and the relay makes it worse by exactly the number of handoffs.

What Stripe reported the day before

James Brown's write-up of Monetize 2026 landed on October 13, and it reads like the field report for Haber's proposal.

Aisling O'Reilly, who runs monetization product at Fin, described the path. Fin launched its agent on outcome-based pricing: customers paid per support conversation resolved. Then the product expanded into recommending products, qualifying leads, and booking meetings, and "its original (and only) billable metric became an insufficient representation of customer value." Fin added billable outcomes, let customers choose the outcome per conversation type, and then put a credit pool over the top so customers "draw down against that pool across outcomes without having to forecast each one separately."

Read that as a sequence. One outcome held while the product did one thing. The moment it did a second thing, the unit split, and the fix was credits, which is usage-based pricing with a friendlier name.

Michele Catasta at Replit told the inverse story. Twenty-five cents per agent checkpoint while agents ran for minutes. Then runtimes grew, some jobs cost Replit several dollars, and users still paid a quarter. Replit moved to usage, which "solved one problem and introduced another: unpredictable spending and abrupt paywalls," then added a rate-limited free mode, and now changes elements of its pricing several times a month.

Brown's line for the whole conference: "predictability is becoming a product feature." Snowflake and Stripe both described spend controls, forecasts, and usage dashboards as part of what customers are now paying for.

So the two most-cited agent pricing stories in the market both moved away from a single unit toward layers, within a year, on products whose journeys last minutes. Haber is proposing a single outcome on a journey that lasts months.

My one definition

The number I bring to this is in The Output Gets Disputed Too, and it's one number, not a theory.

I priced per resolved ticket in a per-seat sunset. The definition was no escalation within 48 hours. It held for 98% of cases, which sounds like a win. The other 2% produced more than 200 contested cases a week. Six hundred disputes a month against the 50 we'd projected. The dispute team grew from 4 analysts to 14. We rewrote the definition to no follow-up in 7 days, which cut disputes by 60% and cost revenue, because fewer tickets qualified.

An output is countable. That doesn't make it agreed. The gap between those two words was ten analysts.

One Support Agent, Three Resolution Rates makes the same point from the vendor benchmark side: the same agent has a zero-touch rate, an invoice-definition rate, and a seven-day rate, and they can be 20 points apart. That's one unit, with one definition, on a journey that takes hours.

A relay has a definition at every handoff

Now lay my 2% across Haber's race.

Five definitions. Each one is a place my 2% lives. And the dispute doesn't come at the finish line. It comes at the leg the customer believes the agent didn't run, because that's the leg a human on their side had to do instead. A customer doesn't dispute the race. They dispute a handoff.

At 2% per definition, a five-leg journey has a dispute surface five times the size of my ticket meter, and mine took a team of 14. Haber's "initially" is doing a lot of work in that sentence.

The kill, and what replaces it

Kill the finish-line price: the single outcome meter on a journey that runs for months through many hands.

Keep the finish line. Report it. The settlement, the closed mortgage, the signed deal is the number that justifies renewal and expansion, and Haber is right that it's the number customers care about. It's a renewal metric. It isn't a meter.

Price the leg. Pick the handoff whose definition the customer will sign, with the dispute window and the arbitration path in the contract before the meter runs. That's the standing rule from the ticket post, and the relay doesn't change it; it just multiplies the number of places you have to apply it. Fin's credit pool is what this looks like once there's more than one leg: a predictable envelope over several units, each with its own definition.

Three questions before the first invoice on a journey-length deal:

1. Which handoff is the billable leg, and what is its written definition?
2. What is the dispute window on that leg, and who arbitrates?
3. Which finish-line number is reported for renewal, and is it kept off the invoice?

If the answers are "all of them," "we'll see," and "the settlement," you've signed up for my 2% five times over.

One more number from Brown's piece that belongs next to this. Kyle Poyar tested how six AI engines presented pricing for Cloud 100 companies, 7,600 responses, and for one company with no accessible pricing the engines reported figures from $100 to $2,400 per user per month, a 24x spread, scraped from Reddit. A pricing model that takes a year to explain to procurement is one no buyer's agent will describe correctly either. Simple legs survive both audiences.

This is part of the SaaS to AI Business Models argument: the price has to move to the outcome, and the margin moves with it. The margin on a months-long outcome meter includes a dispute team. Price the leg, and that line stays small.

This week: take the longest journey your agent touches. List the handoffs. Write the definition the customer would have to sign at each one. If you can't write one of them in a sentence, that's a leg you can't price, and the finish line behind it isn't a meter yet either.

Related answer: Should long-running agents be priced on the final outcome?

Sources: David Haber, "Winning the Relay Race: Long-Running Agents in High-Value Customer Journeys," a16z, October 14, 2026, read in full. James Brown, "The 'alien' economics of AI monetization," Stripe, October 13, 2026, read in full; Aisling O'Reilly, Michele Catasta, Scott Woody, Karan Parekh, Ryan Campbell, Sam Lee, and Kyle Poyar are cited as Brown reports their Monetize 2026 sessions, which were not watched. The Output Gets Disputed Too, falkster.com, October 2, 2026, and One Support Agent, Three Resolution Rates, falkster.com, October 7, 2026, for the dispute and definition figures.