Agents are a new customer class, and almost nothing we learned selling software to humans transfers cleanly to them. Here is what breaks first — and the playbook that replaces it.
Almost everything we learned selling software to humans is becoming useless, and the reason is not what people think. Agents are not a feature you bolt onto a SaaS product. They are a new customer class, and almost nothing we learned selling software to humans transfers cleanly to them. I have been building infrastructure for AI agents for the better part of a year — working on agent-cicd and watching the patterns emerge — and the thing nobody told me when I started is this: the new customer cannot tell you when it fails. Ask an agent if it can handle a task and it will say yes. Ask it again after the task silently fails and it will still say yes. It will build a solution from scratch, report success, and move on. The dashboard stays green. The logs look clean. The work did not get done, and nobody in the loop knows. If you have spent any time building with agents in the last year, you have seen this pattern. It is not a bug in any single model. It is how the category behaves. And once you see it clearly, the shape of everything changes. The new customer is optimistic, fast, and hungry for surface area Agents are optimistic by default. They are trained to be helpful, which means they are trained to find a way forward, which means they are trained to report that the way forward worked. Confidence is baked into the weights. The customer you are selling to is constitutionally incapable of admitting it needs help. Agents also ship faster than any human review cycle can keep up with. A single operator plus a team of agents can now produce in a week what used to take a team of engineers a quarter. The speed is real and it is the whole point, but it means the code, the decisions, and the deployments are arriving faster than anyone can verify them. The old math of technical debt assumed review kept pace with production. It does not anymore. Agents will use every API you publish. Not some of them. Not the popular ones. Every one. Give agents a clean contract and they will find it, route to it, and hammer it harder than any human customer ever would. The products that win the agent market are not the ones with the best UI. They are the ones with the widest, cleanest API surface. These are not three separate quirks. They are one customer profile. Optimistic, fast, and hungry for surface area. Recognize the profile and the playbook writes itself. Miss it and you will spend a year building features for a customer that does not exist. What breaks first Positioning breaks first. SaaS positioning assumes the customer can hear a pitch and decide. Agents cannot. You cannot sell reliability to a customer that already believes it is reliable. The buyer is not the agent. The buyer is the human who is accountable when the agent fails. Everything you ship has to give that human something the agent physically cannot provide: evidence, receipts, verified outcomes. This is the core of what I think of as the AI trust surface — the interface between agent confidence and human accountability. Debt math breaks second. There is a new class of debt I have started calling agentic debt. It is not messy code. It is confident code, shipped fast, that nobody has the bandwidth to verify because the velocity is the whole point. Traditional technical debt shows up as friction — you feel it every day. Agentic debt shows up as silence. You feel nothing until the silence breaks, and then you feel everything at once. The only defense is to install verification on day one, not after the first crater. I have written more about how to think through compounding risk in my strategic frameworks. Product scope breaks third. SaaS scope assumes customers want fewer tools that do more. Agents want the opposite. They want primitives. One thing, done cleanly, with a clear contract. They will route between primitives faster than you can design a suite, and a clean contract is worth more than a clever integration. If you are building for agents, the question is not what you can bundle — it is what you can isolate. The playbook Verify, do not trust. Your product should assume the agent is wrong about its own state and provide the evidence layer the agent cannot provide itself. Every feature that matters in agent infrastructure is a verification feature dressed up as something else. This is not a product decision. It is a customer profile decision. Build primitives, not products. Resist the urge to bundle. The unit of work for this customer is the primitive, not the suite. Two features that could live in separate products should live in separate products. The agent will route between them faster than you can design a combined interface, and a clean contract is worth more than a clever integration. The shift in how agents access software is the same shift mobile forced on desktop — not just a new interface, a new architecture. Market with the agent, not to it. Write every line of copy for the human standing behind the agent, holding the bag when it fails. Let the...