The problem is not intelligence
We already have capable agents. That is not the problem. The models reason, plan, call tools and produce work that would have been unthinkable three years ago. Put one in production and, more often than not, it is shelved within a quarter. Not because the model stopped working. Because the system around it collapsed the first time something unexpected happened, and nobody could say afterwards which decisions were good and which were not.
That is the pattern. It repeats so consistently that it has stopped surprising anyone who has watched it more than twice.
A third thing
Most of the industry's attention is on two tracks. One is intelligence: which model, how to tune it, how to make the agent smarter. The other is safety: guardrails, alignment, red teams. Both matter. Neither decides whether the agent is still running in six months. What decides that is a third thing, and the industry has only just started to name it.
Durability.
A durable agent keeps its state through failure and resumes where it stopped. It can pause for a person and pick up days later. It leaves a record of every decision that an auditor can read. It lets you replace the model without replacing the workflow. It works alongside other agents without corrupting what they share. Five properties. They are not features you add later. They are the architecture, or they are absent.
We have been here before
Before transactions, every application handled partial failure its own way, and most handled it badly. Money moved once and was recorded twice. A transfer half-completed and stayed that way. Then the database took the problem away from the application and solved it once, underneath, for everyone. Nobody writing a banking system today reasons about half-completed transfers. The guarantee lives in the substrate.
Agents are at the moment before the substrate. Every team is rediscovering, at its own expense, that a model calling tools with no durable state is a demo. The 14% who have reached production scale did not get there by finding a better model.1 They stopped treating orchestration as plumbing and treated it as the system.
The position
Durability is separate from intelligence. You can run a dull durable agent and a brilliant fragile one, and in production you want the first.
Durability is a property of the system, not the model. It cannot be prompted into existence.
Durability is testable. Five properties, each with a question, each with a threshold. Score it.
Durability is achievable now. The execution engines exist and are proven. What is missing is the decision to build on them from the start rather than retrofit after the first outage.
And for the agents that matter most, the ones that write to systems of record, durability means one more thing. When step four of five fails, steps one to three are undone, automatically, and the record shows what was refused. That is the part no platform ships. It is the part we build.
Build the 14%
The model is not your moat. The agent is not your moat. The orchestration that makes them durable is your moat.
Build the 14%.