An agentic development startup called FeltSense spent YC's most recent demo day proving a point nobody wanted proven: they reverse-engineered and rebuilt every startup in the batch, autonomously, inside twenty-four hours. Not prototypes. Working clones. I read about it and asked myself the only question that actually mattered afterward: if that is true, what still protects a software business at all?
Three things, and none of them are code.
The first is presence in the physical world. An agent can clone a web app overnight. It cannot show up at an artisanal mining site, walk the terrain, sit through the meeting a cooperative needed before it would trust anyone with its records. Software that only exists on a screen has no defence against a faster screen. Software tied to field agents, site visits, and hardware deployed somewhere with intermittent connectivity has a defence that has nothing to do with how good the code is.
The second is data nobody else can access, because it was never digitised in the first place. An AI agent can only work with what already exists in some system. Enormous volumes of knowledge, particularly across Africa and the developing world, have never been structured or recorded anywhere at all. Every application we build is a channel through which some of that invisible knowledge enters a digital system for the first time. Nothing can clone data that does not yet exist anywhere to be cloned, and once it does exist, sovereignty controls decide who gets to touch it and on what terms.
The third is the strangest one to say out loud, because it sounds like the opposite of a moat: operating in markets that are genuinely difficult. Politics, institutional distrust, turf wars, bureaucracy, governance that fails as often as it holds. These environments are painful precisely because no amount of technical cleverness substitutes for the slow work of earning trust across stakeholders who have good reasons not to hand it over quickly. Years spent building relationships with mining cooperatives, government agencies, development finance institutions, and civil society organisations is not a cost centre I am eager to escape. It is the thing a competitor, human or AI, cannot shortcut, no matter how fast their agents write code.
These three reinforce each other rather than sitting side by side. Presence earns trust in complex environments. Trust unlocks the right to capture data nobody else can reach. That data makes the platform genuinely more valuable, which draws more clients working in similarly hard sectors, which means more deployments, more trust, more data. The moat was never really about the exchange layer, which plenty of people could plausibly build. It is the whole portfolio of applications that keeps generating data nobody else is positioned to generate.
Here is the part I did not expect to believe as strongly as I now do: AI-assisted coding is not a threat to this position. It is an accelerant. If code generation genuinely does commoditise the act of writing software, then the businesses that survive are the ones whose value never lived in the code to begin with. Ours doesn't. It lives in knowing how to decompose a messy domain problem into something a system can capture, and in the data that decomposition produces once it is deployed somewhere real. If a wider set of partners can build compatible applications faster because building got easier, that accelerates the flywheel rather than threatening it, on one condition: the expertise that makes an app worth building has to be encoded in the tooling itself, not left to whoever happens to be typing.
So: what is the value of software when an agent can clone any app in a day? Not the software. It never was. The value was always what earned the right to build it, and what got captured once it existed. Neither of those clones overnight.