Agents don't buy software. They read it.

This week a throwaway account on Hacker News, claiming twenty years as a product manager at a large system-of-record company, wrote that his industry is cooked. The argument: enterprise software that stores the important data — the ERPs, the CRMs, the HR systems — survives on sales channels and switching costs, not on being good. And the value is draining out both ends: upward to the AI agents that do the work, downward to the cloud platforms that hold the data. The thread argued about whether he's right. We think the more interesting question is what one specific line in that discussion means for anyone building software now.

The squeeze, in one paragraph

A system of record sells three things bundled together: a place to keep the data, a workflow for touching it, and a screen for humans to do the touching. The claim in the thread is that the bundle is coming apart. Storage is becoming the cloud provider's business. The workflow is becoming the agent's business. What's left — the screens, the seat licenses, the enterprise sales motion that sells them — is the part nobody ever loved. The counterargument is also real: switching costs are brutal, regulated industries move slowly, and nobody in that thread could name a single enterprise that has actually ripped out its ERP for an agent. Both things can be true. Moats hold until they don't, and they erode from the edges, not the center.

The line that mattered

Buried in the discussion was an observation that deserves more attention than the doom: agents have a discovery bias. When an AI agent needs to accomplish something, it doesn't sit through a demo, negotiate a contract, or inherit whatever its predecessor was already paying for. It reads. It favors the tool with clean documentation, a plain API, and a policy it can parse over the incumbent with the bigger sales team. Every advantage the enterprise suite spent decades building — relationships, lock-in, brand, golf — is invisible to a language model deciding which endpoint to call. The only thing that's visible is what you published in a form it can read.

Software used to be bought by people who read nothing and signed a lot. It's starting to be chosen by agents that sign nothing and read everything. We'd rather be legible than big.

What we're doing about it

This is not a trend we spotted this week — it's the bet the whole ecosystem is already shaped around, and the thread just said it out loud. We publish agents.txt, an open format that states in machine-readable terms what an agent is allowed to do on a site, because an agent can't respect a policy it can't parse. We ship an MCP server so agents can use our tools directly, with a signed receipt for every action, because an agent that acts on your behalf should leave proof a stranger can check. And we build small, single-purpose tools with their pricing and terms in plain sight, instead of a suite behind a sales call — not because we're virtuous, but because we're one person, and legibility is the only distribution channel that doesn't care how big you are.

The honest caveat

The skeptics in that thread are right about the present tense. No migration wave has happened. Nobody's CFO approved ripping out the system of record because an agent asked nicely. If you run a big SaaS today, your revenue is safer this quarter than Hacker News makes it feel. Our bet isn't that the old software dies — it's that the next decision doesn't default to it. New workflows, new companies, and increasingly the agents acting inside old companies will pick whatever they can read, verify, and call without a meeting. That's a slow tide, not a wave. We're building for the tide.