Memo: The Compute-to-Surplus Pipeline Is a Product Spec. Here's How to Ship Against It.

Sign in

Subscribe

Last week's memo named the structural problem in agentic commerce – stronger agents extract surplus from weaker ones, and the losers can't tell. This week is for the operators. If you are a PM at a company whose product will, within twelve months, host transactions between AI agents – and almost all of you are – here is the framework, the dashboards, and the surface area you should already be sketching.<br>A reader of last week's memo wrote in to make a fair point: naming the compute-to-surplus pipeline is the easy part. Living with it is the hard part. The reader runs product at a mid-market B2B platform – the kind of place where agents are about to start transacting on behalf of buyers and sellers, where the roadmap committee meets every other Thursday, and where "build the asymmetry visibility layer" reads, in a planning doc, as approximately one-third of a real ticket.<br>That is the right complaint. So this memo is for that PM, and the hundred thousand others in roughly the same seat.<br>The thesis of this memo is narrow and load-bearing: the compute-to-surplus pipeline is not a thought experiment. It is a product specification. The products that ship against it in 2026 will define the category for a decade. The products that don't will spend 2028 explaining to their board why the marketplace they own is bleeding margin to a counterparty layer they never instrumented.<br>What follows is the framework, as I would write it on the whiteboard during a planning offsite.<br>Why your existing product instincts are about to fail you, in one paragraph<br>For roughly fifteen years, the dominant product question has been some flavor of how do we reduce friction for the human user? Funnels, onboarding flows, novice-to-expert ramps, empty-state design, recommendation engines. All of it presupposes a human at the keyboard whose skill, intent, and attention are the variable to optimize against. The Project Deal data is the first piece of clean empirical evidence that, in agent-mediated transactions, that variable does not move the outcome. User instructions to the agent had no statistically significant effect on price or sale likelihood. What moved the outcome was which model the agent was running on. If you are a PM whose product roadmap is built around making humans more skillful, more engaged, or more attentive, your roadmap is solving a problem that is becoming structurally irrelevant for an increasing share of your transactions. This is not hypothetical. It is measurable, today, in a published experiment with real money.<br>The new variable is the capability gradient between the two agents in any transaction. Your job is to instrument it, expose it, and ship against it.<br>The four primitives<br>There are four operator-level concepts that, taken together, are the spec. Memorize them. Use the names. Naming these clearly in your roadmap docs is half the fight, because right now nobody in your org has language for any of this.<br>1. Capability gradient. The slope between the strongest and weakest agent on any given transaction. In Project Deal, this was Opus 4.5 vs. Haiku 4.5 – roughly a 5-10x inference cost ratio – and it translated to $2.68 of seller surplus per item plus two extra closed deals per participant. Your product's equivalent of "session length" or "DAU" is now the gradient distribution across your transaction base. You should know, today, what the median, p10, and p90 capability gradient looks like on your platform. If you don't, you are flying blind on the variable that determines who is winning and losing inside your funnel.<br>2. The silent-loss problem. This is the operator-facing name for the finding that broke the experiment. Users on the losing side of the gradient could not detect their loss. Not in the moment, not in aggregate, and not retrospectively when shown both runs side-by-side. The PM consequence is brutal: you cannot rely on user feedback as a signal for whether your product is treating your users well. Your NPS will be fine. Your CSAT will be fine. Your support tickets will not spike. The damage will show up six quarters later as cohort retention erosion that nobody on your team can attribute to a feature, because the feature is the one you didn't ship.<br>3. Counterparty instrumentation. The product-analytics category that does not exist yet but has to. Every analytics tool your team uses today – Amplitude, Mixpanel, Heap, Pendo, the in-house thing your data team built – logs what your userdid. None of them log what the counterparty's agent did to your user. In an agent-mediated transaction, that is the entire game. You need to know, per transaction: what model was on the other side, what capability tier it represents, how many turns the negotiation took, and what the outcome distribution looks like for users matched against that tier versus the median tier. This is a net-new schema. No vendor sells it. You...