Why Agent Failures Keep Getting Blamed on the Model

In surgical handoffs, individual competence is rarely the problem. A study of perioperative teams found that surgeons held 83% of necessary preoperative information. Anesthesiologists held 87%. But only 27% of that knowledge was shared across all primary team members. Each person knew enough. The patient's context didn't survive the handoff.

Agent deployments are breaking in the same place.

A demo works. The same workflow, handed to a different team and pointed at production traffic, breaks within a week. The postmortem lands on a familiar conclusion: the model wasn't capable enough. Upgrade the model, rewrite the prompts, add examples. Try again. Roughly 10% of enterprises successfully scale AI agents beyond pilot. When the other 90% investigate, they reach for capability as the explanation because capability is a legible problem. It has vendors, benchmarks, and a procurement path.

The RAND Corporation's 2024 study of AI implementation failures, drawn from 65 practitioner interviews, identified five root causes. Model capability appeared nowhere on the list. The root causes were environmental: misunderstood problem definition, inadequate data, insufficient deployment infrastructure.

So what actually happened between demo and production? The person who built the demo left the room, and their knowledge left with them. The timing quirks of a particular login flow. The dismissal pattern on a consent modal that behaves differently on Tuesdays. The accumulated "when you see this, do that" understanding built through weeks of iteration. None of it was written down, because none of it looked like knowledge at the time. It looked like building.

The diagnostic tools most teams rely on make this invisible. A survey of 1,340 organizations found that while observability tooling was widely adopted, only 52% ran evaluations to confirm the output was correct. Without that second layer, a context gap and a capability gap produce identical symptoms: wrong output. And at 85% accuracy per step, a ten-step workflow compounds down to roughly 20%. Small context gaps don't stay small. They cascade, and the tools that would distinguish "doesn't know enough" from "wasn't told enough" simply aren't in place.

So teams default to the fix they can see.

Over time, the diagnostic attention migrates toward a different question: does this system know what it's walking into.

Some early efforts to formalize this are emerging. Harvard Business Review recently proposed treating agent integration with the rigor of employee onboarding. The World Economic Forum suggested "agent cards" documenting capabilities before deployment. Whether these become the real thing or the bounded version that lets organizations feel like they've addressed the problem is genuinely unclear. Naming a function and funding it are different acts.

Context transfer resists the interventions organizations are good at. You can't procure it. You can't ship a patch for knowledge that nobody recognized as knowledge until the system broke without it.

But the first step is cheaper than solving the whole problem: stop treating every production failure as a model upgrade opportunity. Sometimes the model is fine. It just never got the handoff.