The infrastructure layer around enterprise AI agents will commoditize itself out of existence

The infrastructure that enterprise teams currently spend months configuring to get AI agents into production may not be a durable business. Jesse Zhang, co-founder of Decagon, a customer service AI company, expects that layer to commoditize within a few years, displaced by agents capable of assembling it themselves on the fly.

That is a pointed bet. Deploying AI in production at enterprise scale today involves substantial bespoke work: routing, memory, evaluation pipelines, integration layers, and the human specialists who configure and maintain them. Much of the services revenue and headcount that has accumulated around enterprise AI rests on the assumption that this complexity is sticky. Zhang’s position is that it is not, and that the agents themselves will eventually absorb the work.

Zhang is careful about the shape of his own certainty. His verbatim framing is explicitly exploratory: “once that gets commoditized because the agents can build that on the fly, that I don’t know, and we’ll figure out in three years from now.” That is not a confident engineering roadmap. It is a founder’s read on where the trajectory points, offered with acknowledged uncertainty about the exact timing and form. Readers should hold it as a directional call with a named horizon, not a firm product announcement. The three-year window functions as a check-in date, not a guarantee.

"Once that gets commoditized because the agents can build that on the fly, that I don't know, and we'll figure out in three years from now." Jesse Zhang

What Zhang is describing is not simply that prices will fall as competition increases, a mundane and safe prediction for any software category. The mechanism he points to is different: the agents themselves become capable of building the infrastructure that currently requires human configuration. If that is directionally correct, the commoditization is not gradual margin compression but a structural shift in what the deployment layer is and who produces it. The distinction matters. Gradual compression leaves room for incumbents to adjust pricing and survive. Structural displacement does not.

The check-in horizon he names is specific enough to be falsifiable. Within a few years, the infrastructure work that today justifies specialist teams should, on his account, be something agents perform autonomously as a matter of routine. The enterprises that have built practices around providing that work, and the vendors that have priced it as a durable service, face a different competitive environment if Zhang’s read proves accurate. They are not competing against lower-cost human labor. They are competing against the capability envelope of the models themselves.

What makes the call worth tracking is the internal consistency it requires. For agents to build deployment infrastructure on the fly, model capability, reliability, and context length must all continue improving at a pace that justifies the assumption. Any plateau in those underlying curves extends the timeline, possibly indefinitely. The prediction is less a forecast about enterprise software economics in isolation and more a conditional bet on the sustained rate of model progress. If the models keep improving on their current trajectory, the infrastructure layer collapses into them. If progress slows or stalls, the specialists stay employed and the current services model persists longer than Zhang expects.

That conditionality does not diminish the claim. It clarifies what to watch. The question is not whether enterprise AI deployment is complicated today. It plainly is. The question is whether that complication is a function of where the models are right now, or a permanent feature of how enterprises adopt software. Zhang’s answer is the former. His three-year window is, in the end, a bet on both model progress and the speed at which agents can turn that progress into self-sufficient infrastructure. Both have to hold for the prediction to land. Neither is guaranteed. That is what makes it a bet worth watching rather than a settled verdict.