The standard story about AI agents is that they're almost smart enough. One more model generation and the demos will hold up in production.
That story is wrong, and it's costing teams a lot of money.
Watch where agent projects actually die. It's almost never the reasoning step. It's:
- Integration. The agent needs to read from a CRM, write to a scheduler, and touch a legacy system with no usable API. Someone has to build that. It is unglamorous, and it is 80% of the work.
- Permissions. The agent can technically do the action. Should it? On whose behalf? With what audit trail and what limits? Nobody scoped this, so the agent either does nothing useful or gets access it shouldn't have.
- State. Real workflows span days and interruptions. Agents are usually built as if the world holds still for the length of one context window. It doesn't.
- Failure handling. What happens on the third retry, the partial write, the ambiguous result? In a demo, nothing. In production, that's Tuesday.
None of these get better when the model gets better. They're systems problems, and they need systems engineering.
The uncomfortable implication
If the hard part is plumbing, then the winning team isn't the one with the best prompts. It's the one that can build reliable software fast — and happens to use models as a component.
That's a very different hiring profile, a different architecture, and a different pitch than most of what's currently being sold as "AI transformation."
It also means the durable advantage is not model access. Everyone has model access. The advantage is in the boring infrastructure around it: clean data boundaries, real permission models, durable execution, observability at every seam.
The model is the cheapest part of your AI product. Design accordingly.
What we do about it
We build the plumbing first and put the model in last. Define the actions the system can take, the guardrails around each one, and the state machine that survives an interruption — then let a model drive it. Boring, and it ships.
The corollary is that a lot of "agent" projects should be a workflow with one model call in the middle. That's not a failure of ambition. It's the version that works on Tuesday.
More on this: AI isn't a feature, it's how we build. Or see what we build.